AI in Optical Engineering Panel Discussion - Ep 10 - Rays and Waves

Steve went and did the first dress change of the day got him mixed it up Hi everyone welcome to the race and waves podcast. I'm Daniel Dumer's cook and I'm joined as always by my brilliant friend and co-host Stephen Jones Hello, and today we're also joined by a whole gang of people. We're doing something special today It's our very first panel discussion and we're going out strong by tackling a topic no less complex and current than AI and Specifically AI in optical engineering There are a few topics that are discussed as widely and with as broad world-changing potential as the AI revolution that we're living through The jury is still out on whether it will be the cataclysmic event that people like the Nouvelle laureate and godfather of AI Geoffrey Hinton is for warning or if it will be a more gradual and peaceful adoption like metas chief scientists Jan Lecun and other AI pro voices are predicting However, it is growing increasingly clear that most so-called knowledge work will be affected somehow by the AI development So today we've assembled a stellar panel of six experts at the bleeding edge of the optics world and We will discuss how artificial intelligence is and will affect our lights and to the industry Thank you so much for being here today to discuss this topic with us. Let's kick off with a round of introductions of our esteemed panel Jenny could you start by telling us a bit about your background and what you think the listener should know about you in relation to this topic Thank you, um, it's pretty exciting to be here. So my name is Jenny atwood I work for the National Research Council of Canada. It's a government research center that's focused on many different scientific topics and The area I work in is astronomy So we have a government mandate to support astronomy within Canada and one of the ways that we do that is By building instruments for the global telescopes. So we have both optical and radio telescopes that we support and that we build Instrumentation for and so my role there is And the optical sciences lead so that's optical engineering and adaptive optics my background is optical engineering and physics I went to Colorado College for measures and physics University of Rochester for masters and optics and then I had a short Stint at KLA ten core in California before coming to the NRC. Great to have you here. Um Craig. Can you Enjoy the round table. Um Craig Olson I'm currently with a company called airtay associates. I had a background in electrical engineering from Georgia tech And then I went and got my PhD at the University of Rochester in the last millennium ironically doing semiconductor lasers, nanolithography, vector beam stuff, nothing to do with lens design and then Moved out to California started working in the telecom boom in the early part of the millennium or century or decade However, you want to look at it doing telecom devices commercial type stuff Then I moved over and did some consumer product stuff We did projection light engines back when rear projection TVs were still in the market Which tells you how how long ago that was did some injection molded optics for a while Looking at making cell phone camera lenses then in 2005 I went to work for company called L3 at the time which turned into L3 Harris So for 19 years I worked airborne defense optics Recognized military systems was a technical fell there for a few years and then exactly one year ago one day I started working in airtay kind of more or less doing the same thing So not a lot of exposure AI during the workplace so far But a lot of exposure to things people have tried over the decade So this is going to be a very interesting discussion. I'm very glad to be part of it. Thanks Ronan do you mind continuing? Not at all. Okay. Hey everyone. I'm a runny and seal. I'm an optical consultant really honored to be here with so many experts optical consultant that just means One day sometime in 2012. I decided to quit my job and call myself a consultant whatever that means right Prior to that, I think I was in the same class as Jenny doing masters in Rochester and then I took a year off in 1998 Doing a co-op at a company called i4 Corporation in near Boston, Massachusetts And returning to finish my masters in the spring of 1999 and after that doing some optical system engineering work in the Boston area for four years and then went to Singapore and was a lens designer there for four years It was then that I discovered that I actually I lived the science and and aberration theory of lens design But I didn't like the work of lens design Yeah, they're just receiving a table of specifications There are almost non-negotiable and they have been to meet that and meet the mtf at the spatial frequency I hated that in fact when I was in college. I didn't want to be a lens designer But having acquired the skill of lens design was very very helpful And after that, I um joined a company called applied by the systems in Singapore Today, it's called Demo Fisher and that was really fun Just developing optical systems for burst detection using the PCR method And after that rejoined a company where I did lens design work But then I became a sort of like a in-house consultant working in project management Production R&D and things like that and it was then that I decided that well Why not just be a consultant so I've been doing that for past 10 over years and in between Manage to squeeze out some time to write some books and papers Which really surprised myself and here I am talking about AI and Wondering whether or not I get to keep my job maybe or not. I don't know. Right. I don't know But yeah, very interested to see what everyone has to say about AI in in his world I agreed and whether Roy and then we'll keep his job this maybe Akil who was best points to answer at the moment Do you mind continuing? I'm certain all of you will keep your jobs I'm Akil. My background is from the University of Rochester for undergrad in optics I immediately left there and went to ZMAX Spent time working on the engineering team working with our customers I think I was the first sales engineer there as well So I worked with a lot of the key accounts which are you know primarily big tech and big defense as you all know And then I'm moving on to the product team where I was really focused on you know What can we do to make this offer better for? Everybody in the industry not just one specific customer, one specific group And so I spent a lot of time really thinking about you know what are the the broader things that ZMAX and an optics studio could do to make like easy for optical designers Eventually we had kind of a successful acquisition and I ended up leaving to go to Amazon I owned kind of the all the on-road tracking of all packages globally So at that point I knew what was going on with every single package in the world Nothing to do with optics move from Amazon into a few different AI startups and Just recently I've founded Praxill which is an AI optical design software And that kind of ties together the the roots of my career with a lot of the new information that I've learned from You know the AI world and also how to kind of scale software at Amazon And so I'm really excited to be part of this group I think I may be the least optically Technically knowledgeable, but very happy to be here We're very happy to have you here and Following on that connection Aaron would you mind going next and introducing yourself? Sure, my name is Aaron Elliott. I'm a principal optical engineer at synopsis Let's see. I started studying segmented systems in graduate school at the University of Arizona I went to ball aerospace where I worked on a variety of giant telescopes space-based telescopes instruments and in particular I got into the wavefront sensing and control team for the James Webb Space Telescope Building the software needed to align the segmented mirror segments, which was really a fantastic experience Then I went to space telescope science institute to follow JWST I didn't follow it all the way to the launch. I abandoned ship in 2015 and started a DMACS And then we were acquired by Ancest which was Fantastic because I am obsessed with optical system engineering and multi physics And really supporting the hardware builds from inside the software world So I primarily think of myself as an optical systems engineer now Yeah, it's all been exciting interesting stuff that keeps me awake and gives me challenges every day Which is mostly what I want Sounds like it I do need to add that I can't comment on our R&D team activities Obviously beyond things that we've made public I'll be commenting on subjects that interest me personally and I should also say that I consider myself a novice when it comes to AI Yeah, that makes total sense. Thank you And regarding expertise I guess that all depends on where you draw the line And the last person on our panel and another Arizona grad is Cory Boone Woo, fair now Everybody so very excited to be here. So I started so currently I'm technical marketing manager for Zygo I started my career in optics. So I studied optical engineering at the University of Arizona Started admin optics as an optical engineer then the plot twist of my career is going into market it And so you know engineers and marketers typically don't really speak the same language really get along super well But I was there to make sure that we were communicating everything accurately working with all that engineers And so now Zygo I'm sharing the technical knowledge of all the skill engineers there as well as trying to clearly communicate all their solutions and all the products and stuff They have going on But so I'm fascinated by AI both in its applications to optics and then you know a lot of my side quests are in optics education So I'm making short video or kids books just to introduce optics to people So trying to improve those workflows and understanding how AI impacts optics So very excited to nerd out about that and learn from everybody on the panel Oh thanks for having me over here I love your videos when we're done Sure listen up, they're very punchy I was going to say yeah You do a great great job of it Yeah, I imagine I'm explaining it to my parents who are musicians or to like a kids You know someone who doesn't have the optics background Cool, so yeah, kicking things off let's start with with you Jenny So you mentioned you know as a leader of a multi-disciplinary team It'd be really keen to hear your insights on any ways that you you know seeing your team currently using any AI Any of the tools that exist out there or any things that they're exploring and if there's anything in there that's kind of surprised you Or so I have both optical engineers and adapted optics scientists And they do really very different things both kind of around optics But one of the things I would just make a distinction about is What kind of AI we're talking about So there is the chatbots and we call those large language model AIs and they're very text-based And we do use those somewhat in a day-to-day environment and they're you know help write emails help me write project management reports Those kinds of things I should say NRC has its own internal large language model chatbot And I think that that's a common theme in companies to want to keep their data within their company And so these chatbots tend to be kind of internal as opposed to the chat GPTs So that's kind of one flavor And we've even been doing it like redoing our lab layouts It generates nice new floor plans and things like that The other kind of AI which I think is probably what we're more focused on here Is machine learning kinds of AI Which deal with large data sets And they find cool things within those large data sets that's useful to us And I would say that our adaptive optics group is the main users of those kinds of things We also within our group have you know mechanical engineers, electrical engineers, software engineers We even have the Canadian Astronomical Data Center And so our software engineers and our data center engineers Are looking at other ways of using these machine learning methods to kind of sort data as well And so within, at least within my group Using the machine learning Our primary focus is on wavefront sensing So essentially we have conventional optical wavefront sensors They provide pixel data And we need to convert that data into a command for a deformable mirror When the light transmits through the atmosphere, it gets distorted And then we have to convert that into commands for this deformable mirror And that the deformable mirror does kind of the opposite of what the atmosphere is doing And that provides very clear images And so the technology behind wavefront sensing is varied There's the long standing Shek Hartman wavefront sensor But there are very new kinds of wavefront sensors as well Like one of our specialties is pyramid wavefront sensors And these sensors take a bit more math to make that conversion into deformable mirror commands And so where AI is working for us Is not so much in the traditional wavefront sensing sense Like a Shek Hartman wavefront sensor is very linear And so when we can do linear math To convert pixels from our wavefront sensor or slopes from our wavefront sensor Into deformable mirror commands AI doesn't win So we're lucky if the AI keeps up But when we get into a nonlinear regime Where say a pyramid wavefront sensor When you're modulating a small area And your sensitivity isn't quite as good because you're modulating a small area That is a linear regime But if we can modulate over a larger area It becomes nonlinear But the sensitivity goes way up So we're much more sensitive But we can't do the math so well And this is where we're using AI to or machine learning To make these nonlinear matches between the input data and the output data Some of the things that we use We use a convolutional neural networks And this is what they call deep learning So you have to provide a bunch of data The AI learns that data and then it makes its inferences One of the things that is maybe surprising One of the things that we struggle with Is these AI's are such black boxes So we give it data And then it provides an answer out And we don't know what it did And so I think we'll get into this later But if we don't really know what to expect It's really hard to have faith that the AI did the right thing Right So that's definitely one of our challenges I don't know how to overcome that I think I probably expected to know a little bit more about what they're actually doing And they still seem very black box I feel like I'm about to throw this in the deep end But before we're warmed up maybe You know, before I joined an engineering software company I completely did not appreciate what an art and science Just verification testing of software is And in some way what Jenny says I completely get up But then I have this other viewpoint which is Scientists are always given black boxes That's all we've got We didn't know how gravity worked And testing the software, you know, you have to bootstrap your way up It is really hard to figure out How do I convince myself that this thing That I trust this algorithm And in some sense with the AI That's not different to me There's a way that it's the same How do we ask the right questions To really understand the answers Douglas Adams would be really proud of me Like designing clever questions is the heart really of engineering software testing And in a way it's the heart of engineering Because eventually you get to the point where You need to know that you can trust this answer At least for all these other simpler cases before you go to the complex cases So there's some way that we're already familiar with this problem And we have some skills here That I think we'll be able to apply to the black box question It's going to take us some time to figure out what those are I agree I 100% agree with you And we are, I would say our team is working through those problems And we've, you know, as an example The pyramid wave front sensors, the nice thing about it Is it has a linear region and a nonlinear region And so we can verify things on the linear region And then adapt to the nonlinear region So like I think you're 100% correct But this is our job Yeah, and it's interesting too because in engineering software you know We always want engineers using our software This sounds bad, but we want them to mistrust us Because we want them to dig into the details Of what it's doing so that they really really know what the results we In that sense AI is not new to us Is there a distinction there Though just with more traditional software You at least in principle can trace back the logic from You know, the inputs to the outputs or If you start with the outputs going back to the inputs And then I guess one of the things that at least how I interpreted Jenny's comment was that because it is a black box Although maybe technically Mathematically it's possible to trace backwards Just the sheer scale of it means it's not at all practical Do you think that introduces any unique challenges Or do you think it just still falls within the subset of what we've already done We're just applying it to a new extent I mean, I think we are definitely going to learn new skills Especially with agentic AI Look, up erifying And I'm not saying that we need to understand everything For sure that's inside the black box You know, we're going to have to verify how much we trust the results In fact, in the nonlinear regimes that Jenny is talking about It won't involve understanding every single detail of what's in there But it's the process of verification that we're going to pursue or If you don't think about agentic AI Then I think that really is an extension of what we've already been doing If you're starting to think about agentic AI and other advanced AI's Stuff that can potentially change its own source code Make its own decisions Then there are definitely some new questions there It could lie to us But I also think the universe has lied to us before too Or at least we think it has because we're not asking the right questions And we get a result we don't understand And we poke into that So even with advanced AI's and agentic AI's There's a way that this is science engineering In some of the same ways we've seen before Just wanted to add a little bit to what Jenny was saying about the AI black box I guess in analogy, me as a z-max user z-max is a black box to me in many ways But the difference is that if I do a ray trace in z-max I can kind of do a first order calculation myself Manually and I could check things that way and gain confidence But if I were using an AI black box I don't know how to do something manually there With human intelligence So maybe that's also a big challenge there Every lens designer at some point in their career has not trusted the software And written their in ray tracing code And then realizing how much the commercial software is worth Because it's doing a lot under the hood So you learn little tips and tricks for how to look under the hood And I think and of course that that's very useful right z-max doesn't give out its core algorithms And I don't think that they should But if you have some underlying fundamental knowledge of how that black box Responds to prompts then then you get more comfort with it And I think we're just not comfortable with not knowing how to probe the system In a way that we get trustable results Yeah as we develop those AI tools It would be incredibly helpful to build in ways for it to provide feedback As to some of the inflection points like decisions it made For example, oh this material allowed it to get past this issue it was facing Which then you have to verify the things it's telling you are correct But maybe that could at least give you starting points of where you start diving into it To poke around verify Cory I agree with you and I think actually there are a couple of AI optic startups that have started Trying to provide that information as they build their algorithms And that's obviously something I will strive for at production as well I do think like Aaron had discussed like there's a different skill set And being able to understand if you're getting good information when you're kind of Kind of getting information out of this black box A lot of the skills that engineers have and have traditionally had Were about getting to an answer because there was no better way of doing it And if we have a world where we're just going to generate a thousand answers Then your skill has to become well how do I make sure that this answer is the one that I want to use And it's the right one And so I think it's just a different mindset and different skill set Rather than being something where we need to shy away from it Just you know going through even our educational process change What are the things that you care about? You wouldn't have someone like even today in a computer science degree You probably aren't teaching assembly because that's just not worth doing anymore So you need to worry about things like okay the technology allows us to get here How do we make sure that where we are here is something that we can actually Validate is working and and make good progress with Yeah I think you're talking about that this is a new tool And in some ways it's not different than other new technology tools That we've gotten along the way We don't do ray tracing by hand in big books anymore But we don't worry that people don't understand ray tracing because they're not doing that Or maybe a little A little bit of my hands ray tracing is okay So moving on from what's that the precipice of today Maybe getting into future movements Are there any domains that you're Craig particularly excited about the potential of Where AI can be applied for optical engineering? This is a tough one because I don't use AI For hardly anything in my past experience It simply hasn't been robust enough for the things I wanted to do As a bit of background to irritate where I am now is primarily a signal Processing and image processing company So that's where our company has has been investing a lot of a lot of funding In both C&N type deep learning machine learning As well as the generative stuff where can we use this? Can we use it to write proposals? Right and ever company has their own internal data set And a large part of any company's problem is historical knowledge Where is all this stuff? I had 19 years at L3 And by the end of the 10 year there I was You know with 19 years of data floating around a server Just to start an analysis that would take me two hours to go find the files So from a simplistic perspective that's where I would see generative AI being useful From a workflow perspective What we do tend to focus on is there's a huge amount of Data generated by a lot of our sensors So airborne remote sensing, right? I mean there's terror flops per second generated by all of the stuff in space coming down So finding varied signals and sparse data is a very daunting task for everyone And I think that's where from a technical perspective I've seen the most engagement at Aritae Most of my day job is designing and building hardware Or was at my previous job and not specifically lens design So we would do lens design and full end-to-end building of systems and proving them out And see how you mentioned my background is in semiconductors I learned all of this stuff in the lens design and optical engineering that I do now I learned all of it on the job through 25 years Which was fascinating, right? And of course during that regime I recall There's a lot of overlap in mathematics When you get down to the aberration theory and diffraction theory And start learning about Applying the stuff that I learned when I was goofing off playing Halo in grad school To the stuff that was actually paying for my kids college And you know after a few years you realize that there's a lot of underlying fundamental connections The physics and math level between a curve piece of glass that's going to burn an ant or something And control theory for an adaptive opticaloscope and so forth And I think I finally realized, right? You don't really know how linked everything is Until you kind of get your own deep learning in at least one field So that's another area where I think using AI, whether generative or CNN Can help augment an expert's efficiency and workflow, right? When I think about how I wind up using these things and I will because we'll have to It's like using teams, you've got no choice I really look at it as how am I going to make my experts more efficient in the workflow Because I've seen a lot of these things going back to the '80s There were the expert system, FAD And that's just when I started learning about the stuff Yes, I did try to program one in Lisp, I didn't work Because I didn't know what I was doing, right? Didn't know the core of what I was trying to program And then there was more of a history with some of the stuff about There was a few papers about automated zoom lens Generation where they would go through and scrape the patent database And try to make a machine learning system in the early '90s When the hardware didn't support, I think Don Dillworth had some papers on this Ron, you sent out some good stuff on that, so thanks One of the processes I used all the time was I used these program called lens view Which was a great program. It was an automated search engine for the patent database Pretty much any international patent They had taken the work to put in there and you could go search it And you know, that's kind of an up-a-process that could be automated Although they did a pretty good job automated yet and making a user friendly for the experts So just kind of going through the years. That's kind of been my exposure to how How I've seen an artificial machine learning or whatever you want to call it Anytime I say, "Hey, I just assume I'm using error quotes because I'm going to get tired of doing it" And it's so many audio anyways So again, I think another area that I think could be useful is Again, along with the theme of making the expert more efficient is When you're designing a lens or an optical system even at the layout level But particularly when you're when you're optimizing it trying to get something You know, you're running into a point where the design stagnates and you need to know how to make it better And there's a lot of intuition and experience in that But there's also a standard longer list of tips and tricks And I think that's an area where efficiency could come into play If there's an aggregation of "Hey, this lens" was stagnated and over thousands and hundreds of thousands of examples of that Now where you get that, I don't know, you would get an AI that would have the ability to incorporate some Intuition rather than just static performance So that was something I thought, you know, that I run into all the time that could be an area I don't know how that would be accomplished The other thing is just eliminating tedium Through the course of any any optical system design and engineering, you generate massive amounts of data And it's not all useful, it needs to be distilled down into a useful format Not only for me to identify, but also to be able to explain it to someone else, right? Generating a presentation material Just thinking about, so I work with a lot of infrared systems So we have to do a lot of ghost analysis and a lot of narcissists stuff And I use Emacs, so that's what I'm familiar with I have used Code 5 and no matter how you slice it You just end up with a massive amount of data that needs to be searched through for anything To be honest So those are a few of the thoughts I had about the day-to-day stuff And then looking larger, not focusing just on the ray tracing and the ray banding And you know, getting the MTF up, but in actually building a system There's been a huge push in at least in defense, but it's across the board I'm sure it's you're seeing it in a commercial for what's called MBSC model-based system engineering And it's attempt to incorporate everything I call it a digital twin of your hardware And the intent is that everything about the system is incorporated digitally So you have performance, you have cost, you have supply chain, you have all this stuff And that's a massive amount of data to incorporate That can be very specific, but it also generates a massive amount of interconnections required Kind of the Google analogy, right? Why did Google's algorithms exceed so much? Because they looked on the network connections rather than the information on the site themselves So that's a thought that I've struggled in our systems engineers I've struggled with is how to incorporate all this disparate type of data Right? I've got data that just, it's not just numbers and spreadsheets It's vastly different types of data, financial, schedule, technical, performance, performance, sensitivities If I tweak something over here in this area of the supply chain, how does it impact performance Five years down the road or something like that I think that that's something where the aggregation of data could be Again, looking through signals in sparse networks Or huge networks with very unknown network connections Then the other thing, so these days I mainly outsource design Once you get too expensive, nobody wants to pay your hourly rate to design lenses anymore And in fact, it's more efficient to have other companies do it So one of the things I'll grapple with is if my designer or vendor says they use AI during the design Or doesn't say they do it, but they're using it anyway They still need to do some rigorous analysis to convince me that they're getting the job done So in some sense, I don't really care Because the real work still needs to be done I think there's a quote about You know, the role will still need plumbers, so we'll still need optical plumbers Then I guess the last, well, one of the other comments I want to make And then I'll wrap up You know, we talked about developing expertise in one field and being able to apply it to another field And I think that's one of the things where I think learning connections can be useful But there's a counter effect to that And I forget whose name is associated with it But if you're an expert in something and you read newspaper article on that topic That you're an expert in, you're going to find all sorts of holes in it Right? The order didn't know anything about it Oh, this is wrong, this is wrong, this is wrong This is all crap, and then you turn the page and read an article about something completely different And you're like, oh yeah, this is great You believe every word of it, right? I feel like that could be a problem with generative AI Where it's been trained on a certain type of field And it's required to extrapolate some law of physics that may not apply Where are the assumptions behind this expertise, right? Every world is full of assumptions So yeah, so I'll, I'll end with that It's interesting the point you made better Greg about like, you know, parallels between Different branches of physics Because I remember, you know, like in-burst design for like Fetonic integrated circuit components like we're doing that A joint simulation as well And I was reading one of the PhD thesis that was I think quite big in that for Fetonic integrated circuits And it made a comment right at the beginning about how You know, this is actually a well-known field in other domains And this is just the first time it's being applied To, you know, phytonics, I mean not the first, but among the first So it's quite an interesting thing Like, you know, it took a while for people to kind of realize those connections And apply a known solution to a known problem To a new domain Whereas potentially, yeah, that could be a low-hang fruit for AI Yeah, the intent is to collect it for people before they forget it, right? I mean, yeah It's important to keep literature refreshed Even if you write a paper on something that's been done before Just refreshing the literature with the more recent reference I think it's still value added to the community I used to edit the spotlight series and do some journal editing And that was very often you see a paper that, oh, this has been done before But slightly new twist and there's a large reference list That refreshes the citation next And that's useful Right Unfortunately less valued in the scientific community today Yeah But Daniel and I were talking the other days And I made the comment that Because, you know, there's always the question about like interpolation Versus extrapolation for these sorts of models And I made the comment that I get my old job You know, we had the super wise and electrical engineer He just knew everything And he was always the go-to But a lot of, you know, the value that you provided Of course, he could come up with creative new solutions But a lot of the times the most useful stuff was just that he had seen something before So he knew the answer it away So there is a lot of value in that interpolation It doesn't all have to be like extrapolation Like co-workers gave me this because every time they asked me a question I'd say "Oh, I have that in the book here somewhere" And which. I have a colleague who, going back to Craig's comment I have a colleague who broke this up in a way that I appreciate Which is The ways that we think that AI and machine learning tools can help us Can roughly be put into four categories Hard data analysis There's as an assistant And with generative AI, there's potentially as a colleague And you know, we're an engineering company We're a software company, so we're obsessed with algorithms And I would say that we've already seen an impact You know, an impact in basically all of these areas And maybe as the. We used to call them "graviards" The people that had all the knowledge But I obviously. Careful Object to that girl I did too As archivers, archivers of knowledge And making that whole Finding data, finding documentation That's a super valuable area too So assistant as in helping you with your day-to-day work But then also archivist All of these areas are different areas Where AI can help us and is already helping us But they're all very different applications And usually different, you know, different algorithms too A really interesting way of breaking it down That makes a lot of sense when you lay it out like that But I hadn't actually considered it in those terms before Yeah, I really appreciated it when I saw it Because it helped me start bending things You know, that I'm just saying AI is really going to help us in the house And also to back up what Craig said Completely agree that a system engineering is really a place Where we are already seeing payback on this Understanding how everything interacts with everything else Usually requires these elaborate databases And reduced order models of how things are connected Or pull up NBSC which connects all the complex software tools together Which is another way to do it But it's all very challenging We're already seeing some advances in those areas with neural maps And helping us use reduced order models And understand where everything interacts And better yet than doing higher level optimizations To find the best system for an application Or candidate best systems for an application Where you're considering everything The optics, the mounting, the thermal, the cost Maybe not quite the political situation yet But I think that's what you're talking about So you can really see how everything's tied together In these reduced order models And do these kinds of analyses and exploration of the system solutions much faster Aaron, you have just perfectly either teed up or preempted The question I was going to follow on with you So maybe if you could just continue on this a bit more Like you mentioned at the beginning how excited you are By these big messy, multi-physics problems And so I was just wondering if you could I know you touched on it there But if you could just talk about that for a bit Like where you're most excited or any specific examples you can kind of point to Yeah, you know aerospace and defense has a tendency to use system engineering To the max more than other industries have And that, that is great because once you learn the art of the system engineering You're like, oh, if wish I could go back and redo everything I worked on When I only saw this piece of it System of the engineers are, you grow into it But what I'm seeing now, and an interesting thing about being an engineering software Is that you tend to get an overview of the industry That you don't get necessarily in other places It's been fascinating What I'm seeing now is that as cost are driven down Torns is in everything get tighter Especially in commercial consumer optics They are needing to learn the same kind of system engineering That we use in NASA and aerospace and defense industries And everyone's building things they're more complex with tighter tithes All of these things are harder The system engineering gets more and more important And these AI tools are coming at a really good time to help us with that These are non-trivial problems So having some computation assistance is great If I could push you like, because you mentioned the four kind of categories of Hard data analysis assistance colleagues as well as archivers Is there a particular one of those that you find is most Immediate, maybe being applied for these like big system level engineering So it's the problems Yeah, these are all machine learning algorithms really For the most part right now We're using the archivist capability inside AncestGPT and finding data It's helping the large language models that are helping nest just with Chief and stuff like, I knew how to do this in visual basic But I don't know what it is in Python Great for stuff like that But the algorithms that are closest to our hearts of course are the Machine learning, neural net, deep learning type algorithms Because we really like, you know, we're at a software company because we like algorithms really And we like the ones that tackle the physics So all of the ones that I would say we're most excited about are Of a hard data type We want to train it to do things We need to feed it good data We want it to figure out better ways to get to the end result A really important part of this for us is The things that make sense to tackle are the things that usually that generate a lot of data But there's a piece of this that I think it's lost a lot which is You need to know whether that data is good or not What's the solidity? And that discussion seems to not happen sometimes Like the problems that we really want to tackle produce a ton of training data But you also know how good that data is Because even for AI, garbage in and garbage out still applies So we are thinking hard and trying things You know that we think make a lot of sense and we're seeing payoffs now And there's definitely low hanging fruit versus things that maybe We don't see perfectly clearly yet that are more advanced in the future But right now there's a ton of low hanging fruit to tackle That's exciting Jenny Yeah, I just wanted to touch, like from an optical engineer and not a software engineer perspective Like I took this question and I kind of tried to break it down into the processes that I go through Because I see this really as being more of an assistant I like the assistant thing a quick anecdote So I think it was during my geometrical optics class Taught by Brian Stone maybe Where he said Z-Max or Code 5 is a tool and when he hit that optimize button You'll get an answer But do you know if that's the right answer? You have to take my class if you want to know if that's the right answer So I like the assistant Z-Max is an assistant I know how to do it but it does it easy for me And so if I broke down what I do as an optical engineer And what would make things easier for me I think that there's different things So like maybe this works for paraxial is I have a set of requirements Can I give you a set of requirements and you can give me 5, 10, 50 starting lens designs That potentially will meet those requirements Or our optimizer, I think our optimizer is actually like the original AI almost But when you get into that global optimization landscape You were left there running for three days Could AI do it in an hour? So those kinds of things tolerancing is always one that I find tedious I like to find another way to make tolerancing easy And I think there's a ton of data there Like you run a Monte Carlo And so the thing I always struggle with is I run 10,000 Monte Carlos But I'm only going to build one How do I reconcile that? So can you give me some sort of insights back on like probabilities For building one when I have 10,000 options And I think that that actually within ZMAX we already have the optimax connection And so if you have the manufacturing processes that you can tie together with that tolerancing Would be one And then at the end when we're doing alignments Is your alignment right? So if you have some sort of optical feedback Reverse engineer my optical design Tell me what I have, what I need to fix Those kinds of things, I mean I generally can read an interferogram or whatever I need To make some adjustments But some of these optical designs are very complex And some of the things I work on are very large So it's not really, it's not like I could put something in front of an interferometer I have basically 3D space measurements And I've got one way for an sensor So I think those kinds of things are all like just going through the workflows And what kinds of things would make life easier for optical engineers I think would be a really good starting point Yeah that's a very interesting perspective that you give their own manufacturing only one device As opposed to like in the vehicle industry for example Where the render is churning out 10,000 cameras a day And then that tolerancing perspective is an entirely different one Yeah Yeah I've had the same high mix low volume problem But the point is I had the discussion with Mark Nicholson several years and many years ago Before I pick studio at a conference and I think you've had Mark on your podcast But I told him like I think of ZMAX as my ID My integrated design environment like visual studio or something And it just made me think of like maybe there's some sort of vibe designing I hate that vibe coding buzzword but You know when you put it in that context where you're managing all the aspects of the engineering life cycle within one integrated design environment I think that's a mindset Where the AI assistant becomes like I said managing tedium Yeah I couldn't I couldn't agree more the assistant part is going is great And it has tons of potential Yeah, could I actually love that you said that it's like your IDE Because I've found that calling peractual cursor for optical engineers Is the biggest way for people to understand exactly what I'm talking about Because I think it most investors and people outside of the optics world Immediately understand what vibe coding is and what would a better IDE looks like Yeah, so one of the things Jenny said is the thing I'm currently most excited about for AI And optical design which is the generating initial starting points for design So if you had a neural network trained on real examples So again garbage and garbage out need a good data set going in of existing designs And the end performance of those Then that way when it starts it may not be running the full ray tracing physics simulations for everything But knowing this bank of starting designs and how that's associated with the final performance Generate the selection of all right here are some very highly likely successful designs you could start from But then you do need the physics intuition and background knowledge to understand Which of those are nonsense and which things are more closely related to what you actually So that and then just getting rid of tedium in general and the things I'm personally most excited about In the space for AI it seems like that should have a bright future ahead So you've touched upon the fact that there are incredibly Complex intertwined multi physics components of this and the fact that there are sort of established Software's that are doing a great job for simulating many of these I'm curious a kill you coming at it from a slightly different perspective Creating a new branch of software's that they're doing similar things What do you see the future of AI in optical engineering to be Yeah, you know I Whenever I think of this I kind of take a step back and think of like what is AI doing generally for software even kind of All of humanity and in kind of the best case. I know we've all heard of the SkyNet worst case scenario But if things go well, what does that look like and for me? I think it's not really like just creating faster software But really unlocking the ability for people who aren't experts or aren't necessarily Given all the opportunities that the successful people today have and being able to kind of give them the tools and the ability to still make an impact One of the stories that I really like is while, you know Hollywood was going through and basically codifying that AI can't be used in Hollywood productions At the same time there were Studios if you want to call them studios of villagers in India who are literally You know living off of multiple dollars a day Who were able to actually use AI to completely produce new tv shows and new films that are now on Indian streaming networks And for me, that's like that's what AI should be doing That's the win and these are people who there's no possible way if they have to go through any kind of Process that we expect out of the world pre AI would have any opportunity to do that They were just nowhere. They don't have the funding. They didn't have the expertise. They didn't have the knowledge But AI gave them the ability to then make something that's actually on TV that people are able to watch now And so when I think about, you know, what can we do for optical engineering? Obviously, I think in the short term there's there's a lot of clear paths of okay, we can use generative AI to kind of help find the right design space and do you know Differentiable ray tracing and using neural networks and machine learning on the tolerancing optimization end to kind of find better Designs but really for me the the end goal is make experts like 10 to 100 times more effective and then allow people who Maybe you know someone's a brilliant artist and can see something from a perspective that all of us having a stand background Just wouldn't see and let them have the ability to actually come in and be meaningfully Contributing without needing to have that level of expertise. And I think you know, that's many years away But that's what I want to happen. And then the short term maybe at least you know new students or adjacent engineers mechanical engineers systems engineers Can at least contribute meaningfully to optical systems without needing to have the years of optical experience that we would have But obviously kind of an expert is always going to have more impact But at least let them be able to contribute in a meaningful way We're going to be fair to capture that as you want to democratize optical design Sure That is the buzzword. Yeah, I heard that that sound right That's an important point though. You need that diversity It's not you need the people to ask the questions that you've just taken for granted So I hardly agree with you on that a kill we We need all those viewpoints participating because we need all those kinds of brains and thought processes to really Take us to new places and of our creativity Yeah, I come to the library um, we were talking a lot about optical design but of course in system engineering I'm thinking about the bigger problem too How do we make sure that the mounting makes sense with structural engineering? Can we a do higher level things that a thermalize this what's the thermal environment like what's the radiate of environment like? We have a lot of things to think about beyond just optical design. I keep trying to convince people that Optical simulation is just this tiny chunk of what we do what we're really doing is building devices Yeah, and actually going off of that to get I guess more specific and but I actually think Maybe the future of optical design could be specific uh like building in this native kind of CAD interoperability and being able to kind of Take into account that we are actually going to have to build something and it's not just An equation for a surface from what I hear. I think zmax is already doing that. I think every upcoming startup is working on that and I think That is another real where AI does give us the ability to say hey look at all these 20 systems. We're going to care about At one time and so yeah, I do hope that we we find a good solution for that within the next couple of years I work with cryogenic systems and you know, there's not a lot of cryogenic data on glass Every melt it seems is different like the cryogenic data that I do have You know, there's slight differences and and so your index tolerance has to encompass the differences between the different melts and all this kind of stuff That seems like an area When Aaron said thermal made me think of it where if we could generate a lot of data on glass glass statistics That that could be fed back into these physics models and Optimization models because one of the things I find is I don't know that I really trust The CTE variations and the index variations when you get down into these cryogenic temperatures So yeah, that would be one one area The in AI might be able to extrapolate a little bit better than you know our basic now. Yeah, I think that would be awesome I follow up question and that maybe in software development there's this growing concern about Heavy reliance on AI generated code and so on for that is it risk of eroding developers understanding of their own systems And that's in turn leads to fragile code basis and issues with debugging and so on Is this then creating a risk that emerges in optical design as well? We're engineers rely Too heavily on AI driven optimization without fully grasping why a systems performs the way it does If you can teach an AI to make fragile code then you can probably also teach the AI to make robust code Yeah, you need How are you validating that code? What are the checks that a human does if we can figure out how to do it We can probably teach an AI how to do it right So yeah, maybe the early versions maybe they would result in something that's that's less robust But I suspect it's going to be the opposite we can do more computation faster If we know what robust means or what a quality design means and does that mean it's lineable does that mean that you've reduced the sensitivity Then for sure we can train an AI to do that too. I think the question boils down to are you planning to build bad AI or good AI You know, are you thinking about it from the higher level and all the testing and alignment and that you know Are you thinking about all these things or are we Just focusing on the design part. I think we're you know We're past the world where people are just optical designers. We're we're building things of necessity where we are all systems engineers now I'm just going back to also what I was saying earlier is you know I think we do evolve to a point where the skill set changes and what what's actually expected to be so You know in the same way that I don't think after you leave your calculus classes You probably aren't doing a lot of calculus by hand any derivative is there anything or integrals But you have software that does that for you and you understand what's the value of it is and so now when you use software to do a bunch of different things together You know, you basically trust. Okay, is the software going to be able to do an integral Correctly and then trust that it's going to be right and you know when to kind of use the right calculations I think the same thing is true for AI and I also think maybe a very realistic example is we look at the world of self driving cars Statistically, I think every self driving car that's actually on the road legally allowed to drive is an order of magnitude or more safer than human drivers Isn't true that the people who are driving those cars are less good at driving that probably but that's okay Because they're able to the the goal that they had was get from point A to point B Safely and and quickly and that's being accomplished So it doesn't really matter that the person driving the car is bad at driving the car as long as the car gets them there Better than the average driver is such a low bar though Yeah, everybody's a worse driver than me, right? I want to fall upon hook a little bit at the point you may learn about we can also train AI to make good code I think I have to challenge a little bit because if you look at code 99% of the training set is junk, right? It's bad code So you're training it on bad code and that's the challenge I'm sure at some point we'll get into training sets and truthing and things like that That's a huge area of look at how much the automotive industry has struggled But I think that's the challenges. How do you make sure that 99% of your training set is isn't junk? Yeah, I kills comment about the self driving cars Just made me think one thing that will need to be very clearly defined is who is accountable for the output from the AI tools Similar with our self driving car if it's the self driving nature that the cause is a fender vendor is it the driver's fault Manufacturer the software is being very crystal clear on all that and then the I think a question that led to this was Is there a risk of people maybe losing the skills or if they rely too heavily on the tools? One thing that we'll need to remain is the ability to be accountable for the AI tools So that does require some Fundamental understanding of the calculus you first learned that lets you make sure the tool is doing its right thing Or some fundamental understanding of the underlying physics and how it work in real system to guide the AI So yeah, making that accountability will be very important as these tools get better Or just if people start understanding less and less if they do end up being a little separated from what's truly happening in software So at the risk of being a bit of a devil's advocate here But we have a quite strong proponents for Automating AI in in optical design query. What's it was your perspective on The education part of optics then is that they're seeing its purpose when you're then letting people rely on more generalized models to do design I certainly don't think so because you need people to to verify things are working correctly You'll still need the skill set you'll need a person who understands optics to guide these things to make sure they're working the real systems So I think the optics education is still extremely important and then just the general optics education of Almost everybody doesn't understand this industry that work in so there's some element of the awareness of optics as well Just a more general public So I think that's still extremely important and then the more technical education He's still needed to make sure people can use these tools correct I just wanted to add to what speak to what I kill is saying about doing calculus after college The more I hear that the more I feel like I'm real dinosaur because I still do a lot of calculus And just about every company I went to as an employee I'm the one with tons of paper and pen scribbling with calculus doing integrals Of course every now and then and using excel or some numerical thing to compute very difficult integrals But I do it because I enjoy it. It helps me think I don't want to grow all having dementia And also I just love doing it right and I've never understood why was it that one would leave that behind Because it's just the intuition gained from deriving the equation is Very helpful but I still feel like a dinosaur now But I think you're what you're talking about is something that helps you think and verify what's happening which To me that's the valuable tool going forward. So we don't need to go back and verify algebra again We're bootstrapping our way forward in terms of verification We don't need to go back and verify retracing across this purpose What we're doing is we're bootstrapping ourselves forward and verification to teach ourselves that we trust The results we're getting from the various simulation tools and again in that sense AI isn't so new We need to we always need to verify a complex result and that skill set will be incredibly valuable with AI tools Agreed. So I'm going through with one of my staff who has Bashers of physics and no real optics background And trying to learn zmax and lens design and optical engineering And there's intuition there is practice. There is a lot of mentoring But what I'm finding is it doesn't replace education and so having a solid Educational background in these areas is really valuable I think it's it's the same way moving forward that we're going to want people who are educated in Doing things with AI And I struggle with it, but I'm kind of old but education is is going to be key to doing that You know agreed on on education still being very important. I wanted to go back a little bit to kind of the Building intuition. I think it's one of the things that we talk a lot about or look at is If people will start losing the intuition that they have today But I think on the other side of that when you're able to explore an entire design space so much faster You can start building an intuition that's not possible today And so I think that's really where you know we go from The skill set and the intuition that people have today might become less because they're just not doing it as much But you're now able to kind of really get you know what happens when you have expert designers who have an intuitive grasp of just Not just kind of what the design space looks like but also what the manufacturability is of all of this and for you know 10 different designs that are solving the same problem in 10 different ways I think that's something that we can unlock given new tools that that right now you know The amount of time would take to go through 10 different designs is just not feasible for any kind of commercial purpose Yeah, Jenny like phrase your sentiment exactly like it real expertise is in knowing what questions to ask, right? So that's exactly the position I find myself in now like I'm there's plenty of employees that can do an analysis or design that are going to be way faster than me But knowing the right questions to ask and challenging the right specs, right? So every time I get a spec, I know it's wrong But I need to go back and figure out why it's wrong and knowing the right questions to ask is really The hard part that I think yeah, I share your misery Misery and opportunity at the same time in a way where all of this is talking about critical thinking skills Hmm absolutely the most fundamental character trait that a scientist or engineer needs Healthy dose of skepticism and then some tools to check it. Do you think an AI could ever be skeptical? Oh, definitely Yeah, I think so, you know, can we teach AI some of those critical thinking skills to check? Does this result make sense? It does seem like they're very much Positivity biased in a way where they are very very prone to agreeing with your initial statement And I think that that's maybe a red flag for the critical thinking point that you're making Jenny and that makes me question Also most of those are obviously large language model-based But whether that is the right avenue for the multi-physics-based questions that we're posting in optical design or optical engineering I will say that a lot of the positivity bias I think is found more in their chat interfaces than when you use their API because a lot of it comes from the prompt That they kind of have hidden for if you're using the chat interface Which I agree is a problem But if you are getting deep into using a lot of the APIs and I think Gemini Well, I'm not gonna say a specific one because they change every week I've found that different ones you'll find that they are actually able to kind of do a better job of understanding truth Unfortunately at different times you kind of have to keep up to date with whatever the newest one is There is also always the question about you know, that's the current model And that might be different from what is the fundamental capabilities of these things to do it with some amount of you know further development It's one thing to ask are they currently Capable of critical thinking and it's another question to ask are they will they ever be capable of credible thinking So we've been having a lot of discussion about what we want AI to be able to do Like how we think we can use it to kind of bootstrap and propel ourselves forward and our design efforts and that sort of thing Ron, I was wondering if you could as an experienced optical engineer if you could walk us through where do you think will be the most challenging places for AI to be able to kind of provide value or what do you think the toughest problems for it will be Okay, very good question In fact after hearing what Eren was saying earlier on about systems engineering and systems design That kind of stuff and where AI can help in that I may have to change my answer now So well because my initial okay, let's start with what I where I think AI will have probably No problems with moving forward. I mean having seen things like peraxial come out And I think it's another company called Opto and also photonium where I see it's going is it's going in a direction Where I think it's going to help me because as I said earlier on I pretty much dislike the grunt work of lens design And I want some tool to actually do that for me and I think that AI can do to that At least for the most basic to intermediate imaging assemblies it can do that from start to finish Given that you've got a table of well-defined specifications I want to feed that to AI and I wanted to maybe even call up a lens design program like z-max or also or whatever Populate the merit function do everything do the torrential analysis under standard conditions And maybe even interface with an AI powered optometanical design tool and spit out lens drawings and housing drawings ready to be given to a manufacturing plant I would love that and I think it can do that where I think it will be really challenging would be Systems integration and by that I mean things like complex optical system design Which is not less design right optical system design is taking all these components to put them together These components may not be optical components. It could be firmware software in biotech It could be chemistry the biology all these things so a multi-modal microscope system is quite complex But even more complex is things like the 30 meter telescope that Jenny is working on or the James Webb Space Telescope The Erin was working on it. That's a very complex monster. It requires so many capable people Working together communicating and seeing things and raising flights. That's really complex How do you get an AI systems design tool to do something like that? And I think the challenges there because the AI that we're currently accustomed to Is an AI that is exists in a computer. It's pretty much like a silo So it's not AI in an Android Walking the halls of a R&D department Interacting with people sitting in meetings and raising flights and looking at issues Right in the real world. So as long as AI is not in an Android it'll have a very hard time Developing complex optical systems and just to kind of maybe like at a final remark to that. That's kind of epilogue Most of us here at one point another were fans of our nerds of Star Trek and as far as I can tell right It starts to get out in the inner in the whole franchise of Star Trek. There's never been A permanent Android captain of a Federation Starship has there never right? So and the writers of Star Trek some of them are actual actual physicists really knowledgeable and creative and they don't and they haven't written a script with permanent Android captain So it brings the question is that is that a reflection of our Understanding of the limitations of AI or is that a reflection of our human arrogance? Right we as much de i we believe in we won't let robots tell us what to do right and even when we create robots that Like the Borg if everyone knows what that is right We destroyed them because they're bad So I think the challenge of AI is as long as it's not in an Android forum Walking the halls of the R&D department interacting with the real world. I think that the complexities in real-world optical system design development and fabricating these Systems as what Aaron was saying right. We're building these things It would continually be a challenge, but in terms of lens design. I think that it's hitting that direction the comments of Star Trek I was actually thinking about that as an analogy last night because The ship's computer is a very good example of what we would like to see as AI Right, they talked to it. It just does tons of number crunching the background But many many times they'll ask the computer to do some sort of analysis And I can think of several episodes where you know the voice says parameters insufficiently explained Right, so it's not an all-powerful AI, which would make the show us interesting But the fact that is kind of what a lot of us have in our head. You're right. I mean I still I still watch the show But that's kind of what I have in my head when I think about Generative AI, but it's infatically not and then I just you mentioned the thing about you know human versus robots and whatnot I just piped up to do and as a good example of how humanity is learned to completely distrust Anything artificial intelligent rights are completely banned throughout the galaxy. So There's a real as a whole another podcast. I'm not qualified for but And I don't know if I hope that the doing future is the one we're heading for At least it's not that if we're in the heart and implant it. Right. Hopefully start track instead Yeah, let's keep the fingers crossed drinking Earl Grey tea. So I if you want the start track assistant I think that's where the idea of providing a robust set of assist tools plus natural language processing comes in And Ronan I disagree just a little bit that these more complex problems AI won't be useful for those because we're already seeing some success You know, we built optical devices now in with a huge amount of complexity and the way that we mostly do it is with the reduced order models usually linear In the form of air budgets where all of the terms that talk to one another and are affected by one thing are already interconnected and We're already seeing good results from using machine learning algorithms to explore those parameter spaces and understand and work at those higher levels The tricky part I think right now is that you still need engineers or the AI to determine like What the relationships are and what as Jenny commented if you know whether you have linear or not linear relationships But we've already seen some successes in those areas. I think there's a lot of potential there and that is why I'm a dinosaur But yeah, but you're right. Yeah, the part that's gonna be tough is you know the new things to come up because Humans we have eyes to the real world and we meet each other in the hallways of our departments and we have these random chats and random creative ideas to come out that the AI It's not listening to unless you wanted to unless you develop an AI system with microphones on your hallway Listening to our conversation and then they speak up and then they go oh, by the way, you're wrong about that. There's a new idea here That might happen. I don't know, but but yeah, generally speaking as you were saying Yeah, interesting though you're thinking we have to get more deliberate about what we feed it or or make sure that it is an isolated Yeah, that's a good point So the flip side of all the hallway conversations because I do think we rely on our hallway conversations tremendously Since COVID it's been this push to get everyone back into the office so that we can have hallway conversations But with our younger staff one of the things I've noticed is that they're not having conversations with their chat AI So it's becoming this kind of colleague they're bouncing ideas off of or they're You know asking for confirmation of something that they've done and Almost loses that hallway chat But it's a different kind of thing. I mean, I'm not saying that shouldn't happen. It's just it's a new aspect of our interactions that's happening Yeah, I certainly have seen that change of people who grew up went through school in early career without you know Post pre-COVID versus post-COVID definitely a bit of a social dynamic change there But then just the the comment was raised about hallway conversations, and you know AI not being involved with those things This might be a little too Too forward thinking but as wearable technology continues to advance maybe people's AI Assistants are actually a part of those hallway conversations in the sense of okay, they can still take in that data Uh and be able to process it as well And then definitely as you go through that the the line between black mirror and star trek. It's a little busier Okay, so they're always listening and you know you can go down a bunch of different rabbit holes there But I certainly could see that becoming more of the thing where everyone has their personal and work AI assistants where wearable technology makes that more part of those other things that happened not just in a meeting or whatever information It seems to me at least a lot of this conversation has been around, you know, how much autonomy or uh, do we give The AI to your point to run you know, it can do these close-box problems very well But you wonder about its ability to do these more challenging multi-faceted problems, and then we do see You know, at least in the news like these agentic AI is where it can make decisions on its own But then back to your point there and you know, if such a critical part of all of this is is validating These sorts of systems and you know certainly within software we can let it make its own choices and we can validate without any Substantial real-world implications. Okay. There's probably a server bill somewhere, but uh, you know Mission critical or life critical things. So I wonder you know is now If anybody knows the examples of this that would be great to to share but Because anyone seen or are interested in exploring using these agentic AI's to kind of make those decisions and give them free reign But putting in the safeguards to check and see how well it does and ideally be able to train it to doing a better job I'm a luddite at heart. So I'm gonna reflexively say no Yeah, yeah, gently agentic AI's scare me a little bit because yes I keep hearing the term and I'm a history buff. I wanted to be a history major instead of an engineer But the pay wasn't as good, but the term luddite is often used incorrectly But I hear the term guardrails used a lot and I don't know what that means So yeah, that's my opinion. So here again, so anyone who's for a horticary ill city. I think I'm in the for it can't I mean I think we can think we can take baby steps forward with agentic AI and as we've talked about I believe that we can start Passing on some of these critical thinking skills that we're talking about think we can test and verify it every step I think there's potential here. We have to be careful with it like any new technology Yeah, I'll clarify a bit when I said, you know, I'm against it The luddite movement back in the early 1800s arose when industrialization was first happening in James Watt and the steam engine And people throughout the term luddite as a derogative term where I hate technology for technology sake new technologies a bad thing When in fact the actual movement is is what we're doing now which is Applying hard questions to things that people don't really have answers for so great a steam engine a steam engine saves you Massive amount of efficiency But the impact to the socio economic the political the global infrastructure in the next 50 years was really what the luddite movement was about And we're kind of in the same boat here. It's it's um as a historical analog I think it's actually pretty appropriate because there's a lot of questions that are valid That we don't really have the answer to or know how to get the answer. So just I'm still a luddite, but I want to clarify a bit So you restarting the the luddite movement in the AI age then is this the announcement for that. I would say that that's already started Yeah, yeah maybe I'm not the poster child Shandy I think I could go either way. I'm not sure I have an opinion on this but one of the challenges I see and I don't know what the strategy is for overcoming it Is the fact that there are the global large language models. I don't know like there's global agentech AI's but the fact that so many companies Or organizations want their own internal AI's of whatever flavor that happens to be I wonder how we overcome That data security issue In creating something that's a bit more global I was just curious like what doesn't a global agentech AI Mean I don't know. Sky net. I don't know Yeah, it's you know do we have open source code that is AI or Is it all proprietary? I don't know I guess and against tricky because I want to talk about agentech AI right now There's so many definitions and for me, I think of it always as like basically an AI that's able to access a tool and do an action But I think maybe that's kind of much too practical of a Definition rather than what could it mean on a larger scale? And so that's kind of where I'm interested in I guess learning what Agentech AI means to everyone here in this discussion To me, just me sky net and I'm scared of that So I'll jump in on that since I'm the defense guy here We used to use I mean Termators one heck of a prophetic movie right and in 22 years in defense the term sky net has come up a lot Right, we look at it as a joke you see what DARPA's looking for you see all the stuff that people are pushing for and here's an example of an agentech AI That has been around for a few years now that really there's no answer to is putting in agentech AI in terms of a gun Right, what does that mean right? There's your rubber hits throat kind of question So an agentech AI books my trip to Naples and creates a hotel that doesn't exist and I show up and there's no tell Okay, that's actually been happening, but yeah, the sky net joke isn't a joke anymore based on where people think this stuff can go So not to get too dark there But isn't the invented hotel really just a problem of it doesn't have a way of assessing the quality of its data Largely, I would say it's that as well as the the comment about always wanting to give a positive response Hey, I can't find a hotel here in your price range. So I'm gonna make one up to keep you happy It's both. I think you're right Aaron and then the issue of teaching him the teaching critical thinking skills there But I think that's large language. We're talking about large language models here exclusively now, right? Machine learning neural lats other deep learning Aren't don't face these kinds of challenges Agreed that's a training set problem, right? I mean, I just remember a long time ago when when automotive LIDAR was being developed You would have automotive LIDAR your hear stories about test drives and test drives and you training your LIDAR Road following algorithm for years on thousands and thousands of our data But then you'll do a test drive in this one country road where there's a little flick of yellow paint on the right And the car will suddenly goes off the rails because it's seeing something that you didn't know was in there So that there's a quantity of data and a quality of data for both generative and deep learning I would say Yeah So I'm curious we've talked a lot about automating design tools with AI Mostly to remove tedium and so on that we've mentioned There's another aspect of it that it risks removing the stuff we enjoy, right? So I'd be curious to hear a bit about what you would be sad if AI removed That's just an unfair question Part where I make I make a random random change in the merit function and had optimized in it all the sudden plummets to zero It happens about one every hundred times, but that is extremely fulfilling At a larger societal level the thing that I'm maybe a little concerned about and this is the same thing about Once people had the internet on their phones is the loss of curiosity about why something works Just the increased access to answers all the time which like I just said that's going to something facing people as technologies Introduced constant, but this maybe accelerates that a bit I would say what drew me to engineering was Solving problems if we lost the ability to solve problems then we're just administrators and that's not quite so fun I have this phrase. I use a lot as an engineer you want to hear the phrase. I don't know why this doesn't work because that's a problem You can fix but what I never want to hear is the phrase. I don't know why this works So all fun aside. I completely support your point Corey. I really fear that there's a huge loss of curiosity coming our way I mean in science and engineering in particular though. I mean we're trained to be skeptics and ask the critical questions I just think that's going to keep becoming more and more important And getting rid of this tedium. I just see it so differently like Professional musicians is a good analogy to me Say you're playing piano you want to forget the mechanics of thinking about I have to push my index finger now to get this No, what you want is all of that to go away so that you can think about the overall shape of the piece and The what you're trying to express with it and the mechanics fade into the background Because there are tools that you've mastered. I mean really mastered and you want some of these tedious things in optical engineering to Fade away so that you can concentrate on the higher level questions And are we going to run out of higher level questions? I don't I don't think so Well, just to ask that a little bit or and so I personally I played the piano and in my experience there are The pianists who are much more technical And they're the pianists who are like me who we don't really know what we're doing. We're just play based on what we hear So every time I hear something I like I'll just play it and it comes out automatically I don't think of the mechanics of the keys and everything And I don't even know how to read chords. I just play you know In optical engineering I'm a very low level guy and I like staying that way But who tells us high level questions but state pretty low level other than doing like for example absolute calculus So um speaking to what Jenny was saying I would I would miss the Very highly intellectual stuff, but then again just zoom out again a little bit If you are creative again, how does AI address the creativity portion? I mean, it's there's probabilistic creativity and then there's DNA based creativity Probabilistic sort of like you take an AI tool that by program it can stochastically pick out any random idea And because it's so random at least one of them has to be very creative and doable DNA based creativity those not like that. It's like I'm a human I take a walk in the park and I see something and I associate that thing with something brand new Like a brand new Harry Potter story or something like that AI has no eyes. It's still still still silo and it's based on what we put into it unless you give it eyes and that's That's scary Anyway All roads seem to end its guy net He says you theoretically are it was the extreme version of everything Yeah, I'm hopeful that there's a more optimistic and to this Yeah, it's just interesting that that's the direction things often go And then it would be curious I don't know if there's people out there that have not seen terminator But like we're of that generation I wonder if the younger people have the same adverse reaction to that gender at the AI Like you do I always do more in your less afraid of it kind of thing Yeah, I almost wonder if it's not even like being afraid of it so much as like Like I think privacy is something that maybe My years in high school were the last years where that really was something that existed And now we see a world where people are very concerned about losing privacy But if you grow up if you were born, I don't know 2010 The world without privacy didn't exist so even the concept of like yeah You can go out in the world and people cannot know what you're doing If that doesn't occur to you at all, you don't feel like you've lost anything And so then it's really hard for you to fight for well, let's have privacy She's like well, what does that mean in my life is actually pretty good And I don't see why that's a downside And I wonder if that's going to be the same thing with AI like 10 years from now Probably We should have got somebody at least one person who was much younger than the rest of the crowd onto this AI panel discussion The plutonium founders are about 22 so Is plutonium a pseudo-competitor of yours, I kill? Yeah They're playing the same space Yes, similar space, they just finished Y-combinator, I think they raised like 2.5 million or something So this might be a good time to make a slight pivot Like here we're talking about you know how the younger generation Is going to whether or not they're going to embrace AI We've also talked before about you know what that means for their ability to learn And before the podcast started a few of us were gushing over the celebrity chore gear It was becoming kind of a LinkedIn famous educator So maybe I can just open the floor to you if you have any kind of comments or thoughts about how Any of this stuff might be particularly useful for Educating new generations of scientists and optical engineers Aw, shucks But yeah, definitely useful But like you're not going to see any Larry laser books or videos for me AI generated at me times You know the human touch I still think is very important But it is very useful in the early stages of let's say it's me making a video Or you want to teach a lesson or something for gathering initial Ideas and things about applications or optical technologies Super useful and then for finding actual sources So I'll maybe use AI tools. Yeah initial brainstorming and then Use it to find the true sources to verify things and make sure you have You know a resource that you can trust So it's useful in that sense and then let's say you want to educate someone about optics Maybe it could be useful to Crash things in a way that you're less in play it or here are the requirements where I have to make sure the students understand these things How can I incorporate fun ideas about x, y, and z into those So then maybe the AI tools can be useful there Yeah, but then there's the in terms of content So videos and everything like that the AI slop wave is Giant especially now Sora other tools now are just creating AI videos and a lot of people don't see well Watermarked and might be confused Yeah, so what this flood of things there is just important to have Clear message and cite your sources and yeah for education It'll be hard to know exactly what to trust because there's so much stuff created that just draws for a million things and some of it can be nonsense So trying to organize all this I think will be a big challenge for educators on especially as people find sources now sources To be purely AI generated and harder to know exactly what to verify So things with any kind of established reputation that sense are super helpful a journal, you know Anything where there is a proven track record that will become increasingly important as the amount of slop just increases Although we are also seeing more more papers get written at least partially with AI So even those gold standard sources may not be the gold standard they once were Yeah, and then hopefully they're using it to reduce TDM and you know be able to prepare things no more quickly But then does on those publications and everyone to in order to preserve their reputation to still have a vetting process maintain some level of quality Yeah, I think on the topic of education especially with AI I think there was a point being made in a previous episode of ours where a lot of these Bleeding edge technologies the industry is doing a lot of the legwork in education You get a fresh grad out of university and they're sort of steeped in the The high tech at the industry setting rather than in academia What do you think is the or should be the symbiotic relation between academia and industry and in this topic? Yeah, any co-collaboration on creating curriculum and figuring that out obviously be increasingly helpful And then just the agreement on the skills that are needed for a success when you go into industry or continuing academia Which like we talked about earlier critical thinking Maybe practice evaluating AI results and understanding how to evaluate this and Determine what is worth acting on and what you know, maybe isn't as applicable Does anyone also have any thoughts on how to yet improve the academia side of Making sure people understand these tools Well, I just want to add that you know, I'm very happy to hear Cory say that We won't be seeing an AI generated Larry laser because yeah, nope Yeah, I'm very happy because the value in Larry lasers is a it's a curry boon creation Yeah, that's why people get it associated with all your very nice demonstrations on LinkedIn and YouTube never there That's where the value is and I think you wrote it and did you illustrate that book as well? Cory, I got a higher one of my best friends to illustrate it, which was really fun. Okay, and even I guess we won't see an AI Illustrated Larry laser. It will be you and your best friend. So that's where the value really is And for me, you know as a Mostly optical engineer, but part artists as well scribbling cartoons every now and then When I see AI generated images, I mean there did look cool, but the feeling is different if I see a real human illustrated art The feeling really is there. It's just like how when I write reports I still don't write using any AI tool. It's an extension of who I am or what I am right? So, um, yeah, I'm just I just want to say I'm glad to hear human generated Larry laser Moving forward Yeah, I personally find it's hard to have you a distinct style and voice and Reputation if you play too heavily into those tools, but then you know devil's advocate I guess you could train the AI generative things on your style on your voice and it could more closely approximate a report in your voice Um, so things will get closer to that direction, but I agree. I think a little human touch is invaluable I'll say that I use chat AI is to write a lot of things and one of the first things that I do is I provide it with one of my writing samples And then I say write in my voice because if I don't it sounds awful It's uses the biggest words I can come up with and it always has this positive spin But what I was going to say about education and transitioning from that academic to industrial environments We have a very robust program where we employ students. So From UVic from other Canadian universities. We almost always have at least two undergraduate co-op students A number of our staff are adjunct and I have masters and PhD students And so giving those people the opportunity To work in a real environment with these tools I think is probably in my mind one of the biggest things that we can do for the academic students to prepare them for the real world Definitely, yeah, pretending the tools don't exist isn't helping anybody Oh, yeah Yeah, another comment based on, you know, the literature as I used to be an editor for when the SPI journals and uh I left there before all of this stuff came into play and I'm very curious now what they're dealing with the whole publishing industry, right? How do you tackle that problem, right? Student hands in a paper, researcher hands in a paper, right? I mean you still got the same problem Have you seen those publications like those like meta studies that show the occur Frequency of certain words that are associated with AI's like just going up exponentially and publish papers But that aren't attributing it to AI definitely have yeah the other one I see is the m-dash Which is one of my favorite pieces of punctuation, but now my writing is AI Think I didn't ask a professor recently about this question about how's he dealing with AI and his students Uh, he and he he happens to be um a graduate from Harvard and he was in Capacito's group doing metal lenses and it's like that Anyway, I asked him so so what are you doing about that and his response was I just there's no real way of Dealing with that besides just pretty more weight on exams. Yes rather than um the homework. So that's there's that You know from the software perspective again develop an algorithm perspective I think in research groups there's quite often there's necessarily a lot of algorithm work within optical research groups and You know as neural net tools and deep learning and all of these other tools become more accessible I think for sure think these research teams will start using Various forms of machine learning many are already, but they're gonna be using various forms of machine learning and in optical research That will continue to grow So moving away again from the LLM switch and kind of almost less interested in the part we're interested in is the machine learning the neural nets The part that can really help us with the algorithms I think is just going to be all positive as we make all of these things more accessible As we give access to more and more computation power you know one thing I am interested in in LLM is the experiment of what happens when you give A field an almost unlimited amount of computation power that's that's an interesting experiment So hopefully a lot of that computation power trickles down all of this becomes easier to access and use And the machine learning You know as we've already talked about the machine learning is a different set of problems than the LLM's And for science and engineering to me that's the one that's a bit more interesting actually I am hoping that in the long term educationally that will get sort of computer vision and VRAR on steroids So that students and engineers can start Simulating working with real hardware faster sooner cheaper, you know schools never have any money But if you could do a VRAR optics training I mean our brains weren't based on images and trying stuff I would love to see some developments there And I think it's not LLMs but this would be computer vision algorithms But are they predicting what you're going to see in a picture statistically next Whereas the LLMs are language focused but I would love to see some computer vision work in providing virtual reality and augmented reality tools for education which I think Must have a ton of potential Ernie's and I think you should add a new an additional title to your LinkedIn profile which is futurist Yeah That's what I was in the technology but that's No you are quite the futurist here There was a fascinating paper and I think it was actually Jeff Bezos that just wrote it this week or last about how you know AI is a boom It'll have a bus period but there's different types of booms in history and there's like the tool of boom of an Amsterdam and the 1600s That doesn't leave anything lasting But then you have like the railroad boom in the 1850s and 60s and that left a lot of lasting infrastructure and his point was Regardless of what happens with generative AI They're building massive amounts of compute infrastructure that are going to be extremely useful for what you just alluded to Yeah, there's a financial impact on all that but from a technology perspective You're going to have all this stuff that's that's going to be available and it's going to be Available for things like Inputing massive amounts of AR of AR stuff and And aggregating all that stuff and I think that's going to be a positive lasting effect from whatever happens So I think that's a very good point. Yeah, and certainly a game changer to have the access to that kind of computation power Then you know when it comes back, you know in the field of optics right I'll field of optics because the light is everywhere I mean the Sun is a light source The full moon is a light source the bulbs on our ceiling are light sources you could use all these Physical things to do your experiments visually and like what Cory is doing doing all these nice demonstrations right to use Virtual reality of mental reality to teach and learn and do experiments Would have to be quite niche like things where you cannot find everyday items of your Experiments on yes Yeah, for me I've the best way to teach somewhere about optics is to go to a classroom be in person having to hands on demos and things like that My whole digital efforts and making all the videos was are how do you scale this up? I can only visit so many classrooms close to me So that was the way okay Hey, you can get millions of people to just see something in action But the AR we are side of things how do you continue to scale that up in a way that is more closely related to that in person experience Which is such a cool idea and you're right. Where are so what would you have them simulate? You could do demos like that where it's seeing z-max and real-life get to the align different optics and see how light would bend through that and teach some of those concepts But then this applies to every field not just optics and for medicine or you know the trades just learning how to do something virtual It's just such a cool idea. We're seeing the first steps toward this Some CAD softwares have started to work on at least VR Three-dimensional visualization where you can turn things and you can imagine with the computation power behind it You can imagine extending this a lot reaching more people training people sooner. You know when you can't afford hardware So clearly we have touched upon a vast amount of topics here today Rapping up. I'd like to hear what's the most important thing we haven't talked about here today Well, I don't know but I don't know if this is important But something I'm really interested in is we touched on it But didn't talk about it the fear of intelligences that are not like our own Which I think is showing up in play in in our society right now as we think about AI Uh-huh. It's a super interesting point because I think it's shown up in this Recording as well because one thing I notice like if we go back to the way that you categorized it air And you know there is the hard date and houses problems and everybody seems on board with that There is the archivers everybody's on board with that and then there's the assistants and everybody super on board with that But then as soon as we gave it any freedom Like hypothetically everyone was like oh no, no, no, no, no, okay. I'm being a bit glib about it But it's interesting that we do have that gut reaction to I want this to be in a closed box. Yeah. I don't have anything genius to fall on from that But yeah, the AI does not apply to AIs I guess I'll throw out one thing that I work in defense So I almost never use this word anyway cost There's a cost to this stuff, right? Maybe we're not bearing the cost, but it's coming from somewhere And that's I'd probably be on the scope of what we can attempt to accomplish as engineers because The little integral sign with the line through it I don't know what to do with that you know dollars and cents make I don't know how they work But I think that would be an important you know this infrastructure cost billions And every time you make a funny cartoon in a design review Using generative AI you're you're churning a lot of watts somewhere Both energy cost and financial cost so this stuff sounds great. It's free Not only free because Aaron you guys are going to put all this new stuff in Z-Mex for free, right? Don't answer that. I hope not So I think that's an important thing that we have it covered annual um Well, I was at e-cock the European conference on optical communications Like a week or whatever and there's a talk and I can't remember the dates, but it was something like By 2030 they expect 20 or 30 percent of the world's total power generation to be used Exclusively for like AI training and inferencing and the reason they brought that up in the conference on optical communications Was that it was the presence or location of power sources that dictated how they added to network these things rather than You know just how you could do it if that wasn't the limiting factor But just that number is kind of Wasn't that long ago that everyone was so concerned about making sure we were reducing our energy So that's just gone out the window like that doesn't get talked about anymore. I think building nuclear plants. Yeah Kind of for but I think both we're building nuclear plants and also I don't there's a company in Washington called Helion Which actually the contract with Microsoft to have a functional fusion plant powering some of their data centers by 2028 So I think it is true that the energy costs are going up, but I think in the future Yeah It's really cool. I went to a conference that at Microsoft and they were They were there and they were presenting and I was talking to the CEO and he's like yeah, we have functional prototypes and We're going to be ready by 2028 and I was like okay. We'll see But I think because we have so much more demand for power Maybe one of the upsides is we will Spur a lot of innovation in the in the field of power and hopefully have better ways of powering all these data centers That aren't based on the current methods that we have because those won't be sustainable Another part of that is it's easy for us to I think most of the time worry about the power cost and maybe computational cost of the most Recent model of AI, but if we look at like right now on most of our cell phones We're now able to get like what the equivalent of 3.5 was just hosted locally Because of the you know the the breakters that we've had in the last two years And so if it's true that you know we can always have two years previous Hosted locally on local computers. I think that means that you know There's a lot of capabilities that don't require the newest model in the world If all you're doing is using it to write essays you don't need to have like You know whatever the latest version is you can use you know for is fine And I think there's gonna be a like we'll keep reaching points where it's like okay this version of AI is good enough to solve this problem forever And we don't need to kind of use the latest version or use the energy costs or the computational cost until we get to kind of a much harder problem Another thing that that we haven't or at least I haven't felt like we've touched on as actual metrics for efficiency We're all engineers and we have jobs. We want to measure our time in hours or did cost or You know whatever comes up and I know a key on your site you throw some efficiency numbers. I could be this many times better Um the traditional process So both with an optics and the world at large measuring efficiency and there's many different ways to measure it I think that's another area where he could throw some discussion points and get ideas about what is this working or not right How do I know that this is working or not? Well automotive light art. Yeah, my card did not go off the road successfully Okay, check that works. I got a lens design Concepted optimized and set to the shop for build in three weeks less. That's kind of a no-brainer You know like I mentioned at the beginning it used to take me two hours when I started a project I forgot where my files are that's clearly measurable But I think there's aspects of how we how we know we're getting it right and how we know we're getting it wrong I mean we didn't touch on it. I think that's another question at large. It feels like humans could do that better than AI Well, I mean it goes back to the question of Always trying to quantify things right should we be quantifying things like aren't quiteifiable And it's just like in financial engineering, you know looking at market indices and things like that Do a lot of physicists who have gone to Wall Street to try to model the financial world Which is not modulable based on what I see but just using a gut to measure something might be the way to go The other love that I took the tip I don't I don't even I don't even think that's a love I take I think I think that's where having the human in the loop is always going to be really valuable I had one example I don't remember where I heard it But it was like okay if you designed an AI to build like the perfect transportation system And you're going from point A to point B and there's a part in the middle It's going to say go through that part. It's just like a patch of green grass and maybe that part has a huge You know immense historical value the human needs to make the decision like is it worth going through there? Maybe maybe it is maybe it isn't but that's not something Yeah, I can tell you like what's the value of of tearing down a historic park in order to get a train to its destination faster What he said yes, I think it's a prioritization. Who sets the priorities? Yeah, that's another good question right that comes out of this Who is yeah, yeah, who is setting the priorities for these things? It's the same thought process and importance of if you're setting a metrics for an organization So we're going to measure everything by time to mark it. Okay. Well, what is it sacrificed as you optimize to that? So you need to understand a big picture to okay that but also have a metric around quality and here's how we measure that Say that's that larger picture and defying those things which you could put into a model once if you truly understand What those parameters are yeah, and going back to optics, you know, we're already grappling with that So is this designed better than another one is in some sense? It has a some amount of subjectivity in it, right? Depends what you care about what your priorities are But but none of that is Magic we can make some decisions about how do we measure efficiency? How do we? How do we find a system that's better than another system and On my end personally there's also this sense of my own personal selfishness because I may not want to train AI Right, I don't want it to think like me. I may challenge AI. I may say okay Do your job right and see what you can do without having me let you be me let me be me let you be you Something that because it seems like we have to train like just like I think earlier on Jenny also said that In order to get AI to write something for us in our voice. We can kind of upload a sample I feel very hesitant to Give anything that I that I have To an AI to use I kind of want its intelligence to be real intelligence like come up with your own personality But then again, it's kind of I don't know Jenny So to summarize the reason wave so AI panel discussion on uh, it all comes back to skydot So I think we had a I at least I really enjoyed the discussion. I really appreciate everyone Basically, I just wanted to say thank you everyone from coming I really enjoyed the the discussion that we had we covered quite a bit from you know, how we can class by different types of AIs and their purpose how much hesitancy we have towards where we think they should be applied Where we're excited about their then going where we're maybe a bit more fearful of them Crossing the bridge and becoming sky net, but anyway just uh from the raisin waves podcast Thank you to everyone for joining I've had a boss. Likewise. Thank you so much. Thank you so much. Thank you so much. Thank you. Great for a lot of time. I speak to all of you. All right. Thank you everyone