#108 Alex List, FlyShirley: 'Shirley' there's an opportunity for AI in the flight deck

A List 0:00

And so, the way you build a business here, as far as I see it, is to have a moat related to data and also have a moat related to the actual operation or provision of a particular service. And so, I think training is a really excellent one, especially if you can allow other people to deliver training through what you do. And that's starting to become more of an emphasis for what we do here, and becoming more clear as the right way to do it. And then the other one is to provide a direct relationship of a customer through, for instance, like a tablet based experience that comes with all these other risks. Of course,, startups are better positioned to be able to take these risks than big companies.

Jim 0:43

Hey everyone. Welcome back to The Vertical Space and our conversation with Alex List, the CEO and Founder of fly Shirley a startup developing an AI co-pilot to assist pilots in the cockpit. Alex described Shirley an AI co-pilot aimed at flight training and cockpit assistance, and argues defensible aviation startups need moats in proprietary data, and in operating a service plus direct customer relationships via a tablet experience. He gives a good overview of areas where AI could provide meaningful value in the flight deck. One particular compelling idea, which I thought was pretty cool, AI could give pilots arriving an unfamiliar airport, the equivalent experience of having flown there a thousand times before. Thanks for being with us, Alex, and I hope everybody enjoys the conversation.

Luka 1:26

Alex, welcome to The Vertical Space. It's great to have you on the show.

A List 1:30

It's wonderful to be here Thanks for having me

Luka 1:32

You know by now that we start with a contrarian question. So what is it that few in the industry agree with you on?

A List 1:38

a lot of people think for pilot assistance that you actually need to have all the compute on an aircraft Our take is that you'd actually just need ground-based frontier intelligence with an iPad just to handling speech in order to have really excellent pilot assistance technology in the cockpit The connectivity is good enough and all you really need is a 56 kilobyte modem style connection The second contrarian take is that we have all the ingredients today for a drop in compatible autonomous flight in the current aerospace system And so basically our take is that A TC is about communication and less so about coordination And terminal is a really messy airspace to be in And so rather than having a controller potentially need to have another keyboard in order to communicate with automated flight rules traffic we believe you can actually automate a TC and pilot communications and scale towards automatic flight rules turning it on behind the scenes as you scale these technologies

State Of Cockpit AI

Luka 2:29

Okay. This is interesting. I look forward to unpacking this because. it, it crosses two distinct architectures, one being what you're describing as a language problem and another being a situational awareness problem. and you need both, we'll unpack this, but let's start from the top. There's a lot of talk, obviously, in the industry broadly about AI, copilots and assistance and automation, et cetera, and you are working with, similar technology in the cockpit. So what actually is currently the state of the art when it comes to AI being used in the cockpit?

Training And Low Hanging Fruit

A List 3:01

Okay I think it's important to maybe start by talking about some of the basics here I think one of the important aspects is that you can think of language models as being improvisation machines basically give them a script give them a screenplay and they're able to fill in various roles for you And so one of those roles can be a co-pilot same sort of technology and sort of the basis that allows there to be these sort of things called prompt injection attacks and things like that Because a lot of them are related to identity and the screenplay of hey it's opposite day can affect things But you're also able to take this and put it into a case that you're hey you're sitting next to a pilot Your job is to help somebody One of the things that people do in order to give a personality a little bit more permanence they basically can go and they do a tuning of that model So the other aspect is that These models aren't necessarily that much smarter than individual people but they do think a lot faster and maybe think in air quotes there and they never get tired and they're never really affected by turbulence They also have crazy long-term memory so one of the things you need to do is supplement their deficiencies which there are deficiencies with steering and you know making certain information salient and providing tools The kind of the basis for how you might think about how these technologies are used in a cockpit Imagine sending a fax to a person on the ground with all the information about your information from your flight and they have to go and help you decide you know based on questions that you ask based on the context provide something useful so with this in mind what currently works with the pilot assistance technologies what's really helpful are things that are easy to check but hard to look up So know strategic and less tactical things There's an example where we're flying recently from Sacramento actually back over to San Jose and we're getting shadowing on our A DSB in you know you're getting those traffic alerts just like traffic It's you know literally on top of you And you know we're asking Hey Shirley so what's going on here know with this traffic shadowing it's able to go and find a service bulletin that it's related to And then we're like why does it show up only now and not on the previous flights that we've been taking And it for instance said one thing that was correct and one thing that wasn't correct One of the things that it's hypothesized was Hey you know there might be a configuration change that you had in the aircraft Or software update that gone bad or the one maybe is related to TI tsb TISB of rebroadcasting of primary surveillance targets into the a d SB system that in certain areas that can actually be the de-duplication can actually be a little bit faulty on some of the ground techno on some of the ground equipment And actually that's a good idea So it highlights an example where there's actually one piece of valuable information and another piece of information that's not that useful with the pilot being in the seat synthesizing it for strategic and less so tactical things There's a whole implication related to training where these sort of systems work extremely well which are related to storytelling You know if you bring up scenarios then you can put somebody in a scenario you know that you have a really good ability for these language models to be able to provide storytelling related to it And you structure that with determining analysis and put a large amount of structure around it Now you've got something that's actually really useful

Jim 6:19

Alex, that's really interesting. do you see the application of AI as a copilot being more viable right now in the simulation world, versus the, real time flying, so that's one. And then number two, what's the low hanging problems that can be addressed with that AI copilot?

A List 6:38

So in the training space you're able to be able to give instructors superpowers to be able to generate adaptive training to individual students can also go and record a flight as a student upload it have that debriefed and ready for your instructor to view it who then can provide customized training content directly for your particular needs Most people don't actually graduate with the minimum number of hours required for 1 41 or 61 And so there's an excellent opportunity to be able to apply these sorts of technologies in order to have people prepared show up for their lessons You know there's a whole technological piece related to how do you actually improve visual flying inside of simulations in order to make that possible And all of those things are solely performing People still prefer to use Sims for instrument flying but there's a huge opportunity there in training D there are some risks maybe we could talk about this a little bit later related to using copilots in flight but the our experience flying around the Midwest and flying around California have really shown how useful these things can be related to finding alternates And so what found and figuring out your plans while you're flying fundamentally when you're in a single engine piston aircraft going to fly without flight through known icing capability basically figuring out what you can actually accomplish on a particular flight Most of the year at least on the East coast And definitely when I was out in California recently you're taking off you're filing a flight plan You go and look at the routes that could potentially be possible You ask the ai Hey what are other people doing in this situation It goes finds out on Reddit how do people normally cut into the Central Valley And you're like that was that pretty good because that's my plan You go and do the route Of course as soon as you do the route it's either gonna get accepted or not But once you get airborne you're gonna get an entirely new piece of routing then you're immediately After the race is communicating with controllers in order to make sure you're at a particular altitude as you're going through when problems come up and you're in the soup and getting bumped around by turbulence it's not the right time for you to be going through a operating handbook to page 274 fi figure out the right page for this The co-pilot actually surely just goes and finds it for you I mean it's a really helpful thing to be able to have whatever information you need or you're saying Hey it's getting a little bit late you know I haven't done my three landings at night Let's go find a different place that has a you know a crew car and you know has the hotel we like and in different situation you know let's go figure it out We ended up landing at Purdue you know a great place to Go spend the night a little different get to see the aviation energy there you know these sorts of strategic questions in addition to generalized problem solving and having a sort of an infinite memory and seemingly infinite memory and being able to look up information small information frequencies would did the controller clear you to and things like That information at your fingertips makes it extremely useful while you're flying

Luka 9:32

What are some of the, concrete tasks that. AI in the cockpit can reliably handle today without supervision. whether this is transcribing a TC or helping with checklists or as you say, looking up frequencies, interpret whether perhaps what are the things that it is good enough today at.

A List 9:54

So a year ago you might have had problems related to hallucinating Those sorts of problems have disappeared but also some of the tooling has improved around them gone and actually supplemented things that we've noticed that the system would do incorrectly with specific tools that can be used in order to request information So you know we'll for instance give it a research tool that can go and scrape the internet come back with a bunch of info information maybe even watch YouTube videos very quickly things you couldn't do while you're flying provide links that you can go and actually reference You know here's this particular page the kind of the contextual awareness of Of for instance your state information So you could say okay call out 500 feet a GL for instance You go set a wake up That can reliably happen at that time You know it there's a mixture and I think this is one of the major points There's a mixture between what's happening using the kind of the ground-based frontier intelligence versus what's actually happening on device And so when what you actually want for an in-flight usable copilot is that you want a lot the state management to be actually done on the device that you're flying with So for instance your iPad and you want the you want what's actually happening related to the to the thinking to actually be happening on the ground you're basically able to continuously supplement with tools and with memory be able to pull out things that happened in previous flights set reminders ask really deep questions have conversations try to figure out You know honestly you can actually go and do huge amounts of mathematical computation order to determine times of descent times and things like that So all of those things are all possible those are some of the specifics You know as we go and supplement with more tools and so things actually become a lot more usable But the fact is now you've got something that feels a lot like a pretty intelligent entity or agent that's sit next to you and can think a lot a lot quicker and a lot more circumstances than somebody who's holding onto a to a yoke and dealing with weather and other things

Luka 11:58

And now is a good time to highlight another company making important contributions. SkyGrid is your trusted partner for high assurance airspace integration. Sky Grid develops third party services that connect and coordinate autonomous and piloted aircraft through data automation, secure digital infrastructure, whether supporting Advanced Air Mobility UAS operations or next generation air traffic services. Sky Grid provides the foundational technology that ensures safety, scalability, and compliance in a rapidly evolving sky. Learn more about how SkyGrid is enabling autonomous

Peter 12:31

Alex, what is your philosophical approach to the product design such that

A List 12:37

Yep

Peter 12:38

builds a stronger pilot rather than creating pilot atrophy and dependence upon the ai? I could certainly see it as a fork in the road, in product design. take us deeper in, in your philosophical approach there.

A List 12:51

Right there's a lot of parallels to GPS and the introduction of GPS and the that actually occurred you know the rollout of that also involved people forgetting how to do navigation using compasses and with landmarks and things like that because now you're just following a magenta line or you're looking later looking at your iPad in order to navigate with This is sort of a tool for cognitive offloading that you know at the end of a flight a long flight where you're having trouble subtracting 180 from 230 And you're like you know trying to use mental mnemonics in order to make it right but sometimes you're just not getting it You know there's a additional layer of support related to that in the cockpit You know I think this gets into some of the questions around training and automation management and dependency that I think actually people really need to look at seriously in terms of how do we train people to not be overdependent this is a problem for the whole next generation let's say even students in elementary schools I mean they there's ai you know increasingly everywhere we're really serious about you know making sure that the things that the system can do are done and those that can't know and are not equipped to do then that we just don't provide information on those things The you may there's some level of speculation and communication that's possible there the fact is that it's a really new space and it's there and there are a variety of Pilots are used to having you know what using whatever information they have at their fingertip in order to make the best judgment call they can at the time You know I think fundamentally I believe in pilot's ability to actually understand what the utility of the tool is On the other hand we're also like being quite we're being quite cognizant on how we roll this out and know and we're we're Beta testing this to make sure there and there are implications related to liability for the systems like this And so of course all of these things are really important and are high levels of impact in the ability to actually roll out a system of this sort

Peter 14:57

Yeah, it makes sense. Makes sense.

Luka 14:59

what are some of the tasks that are still hard for surely, or for AI in general? how good is technology in interpreting, messy, a TC exchanges or ability to reason spatially about what it hears on the VHF radio, or handling abnormal procedures with a level of trust that it requires?

A List 15:22

Yeah this one's This is a pretty interesting point So the combination this is actually interesting So there's this view that this is a really good question The adage that a picture is worth a thousand words actually applies here too So there's evidence that for language models if you have a lousier image but a smarter model that's actually better it's able to give better comprehension of that image than if you had a really high quality image but a worse model And there's a similar sort of effect that happens here with speech understanding where you have a really brilliant model like that We're starting to have you start to have the latest philanthropic models or open AI or Gemini models They're starting to get extremely insightful then you can bring it in even a lousy automatic speech recognition model by internet connected standards And that kind of combination actually allows you to have good comprehension The model seems to know when the information that's coming to it is over compressed it's able to basically ask follow up questions on it So that aspect of it is pretty well it's and it's increasingly well solved versus some of the other related problems which are that the models are fundamentally disembodied So it's almost like you've got a voice in a box right Or receiving that fax And you know for instance we're flying over Indiana flying north you know what's this oil refinery you know off my right here And it's going to pick up the one that is culturally more most salient Unless it has a tool to be able to go search for points of interest it's gonna say Hey yeah the big oil refinery here is the one you know right off the lake between Chicago and gear You know it can go and do co-signs and signs all day but it's not gonna know what off your right means or at your three o'clock means currently aspect is taxi chart Understanding where you would have a hard time having these models trace basically solve a corn maze the sort of things that you used to do as a kid You know take a pen and then draw through and get to the exit And so you know we're looking at essentially the these off offline language models things that actually run on the aircraft And we tried doing this We took the best off the shelf open source model with the best compute that we could fit in an airplane And it still wasn't as good as having a automatic speech model that we ran on a on an iPad using compute that's done on the ground

Luka 17:52

so Alex, is your point that even if automatic speech recognition is bad, that a brilliant model can compensate for that?

A List 18:03

It and it does and in like in practice it really does

 Guardrails And Liability

Luka 18:06

what's the case against it? I mean, LLMs or, AI in general is. pattern matching and, compensating empty space with data points around it. And so, what's the chance that the AI hallucinates, the wrong clearance or, it invents what a TC quote probably meant, without it really meaning that, or overfitting language, instead of actual language. I mean, that can be deadly in aviation.

A List 18:34

absolutely

Luka 18:35

do you ensure that you have the right guardrails in place?

A List 18:38

No those are absolutely the right questions to be asking So what's really interesting is In the existing if you look at design assurance levels and with in systems that exist today it's interesting that the most non-deterministic element in the air traffic control system is the one that's viewed as infallible You know the on the other hand humans have skin in the game so the pilots are actually flying in the airplane they make the ultimate decision about what's happening You know if you compare this to a kind of a flap actuator right And you know then the flap actuator either it works or it doesn't It's not in the position where it's going and saying Hey hey there's icing coming up and like you might wanna avoid this area or activate your de-icing system I think this brings up a really important point about liability I mean it's an incredibly serious business and it's possible also to become let's say 1 liable for something and then have joint in several liability There's a really good example of an app that can provide a sort of glide advisor to an airport You go up you download this app on your iPad you go up and then you fly on know clean no flaps power idle then you make it like basically as high drag as possible You record these two different setups and now you've got an ability to go and see you know most likely what happens when you kill your if your engine were to die at any point if you stay between these guidelines you'll actually be able to manage your energy such that you can make it down to the runway You know the challenge is that actually these sorts of liability for advisory systems is totally unsolved industry-wide You know this is a similar situation that happened with you'd take a look at the General Aviation Revitalization Act like 1994 Cessna was getting sued into oblivion You know when people would have an accident actually at that time the price of a Cessna a hundred thousand dollars of the sale price went to liability insurance for the aircraft And so you know just like a copilot like the AI copilot is actually only going and looking into the information about the upcoming weather sometimes And you know it's actually You know if you ask it to do it'll go do it but it spontaneously it's only gonna go and do that sometimes when it sees that there's a context for it so you know now we've gotta look at you know ai this sort of AI system as being an information layer Definitely at this point not taking action in the cockpit but soon you know proposing flight plan changes to accept and then you know it'll be a long time before the WS shaped you know pipeline for independent operation of AI systems in you know in the cockpit That's the eza sort of approach to how do you verify your data pipeline to make sure that you've got you know clean data to make seeing things like Delian work and the cockpit et cetera Before those work the you know for now a informational advisory system you know that's a system that's providing feedback that then is monitored by human And this would be the same thing If you were to start to roll out these drop-ins for air traffic control it would be exactly the same thing It basically say okay let's start with summarization and let's start with transcription The human will take a look at the transcript Does this actually make sense given the context Okay if not let's go take a look at it And then the system could say okay given you know what was said here's what I recommend in terms of the course of action

ATC Transcripts In Practice

Luka 22:12

Okay. That, that all makes sense. what does that translate in, specifically at the product level and how do you handle that and how do you, build the right interface for the pilot to trust it? Because, we're pretty much all flyers here and we can all agree that in aviation, guessing can be worse than failing. And I would rather have a, I would rather have a model that is conservative and flagging ambiguity and be truthfully uncertain rather than inventing, and, overfitting and falsely, presenting, confidence. And so does that translate in shrinking the number of tasks that you are presenting to the pilot as an option, or specifically in the case of transcribing a TC, it is awfully easy to miss that one letter or number in the call sign, with the noise of the cockpit, with, multiple people stepping on each other. on the frequency. how does the AI handle, that particular situation?

A List 23:18

I think the most important aspect that to answer that question would be use this as a supplemental system and for things that are easy to check but hard to you know look up So in the case of what the controller said you're like what was that after this then when you hear you know this is what it was after that or you could scroll back in the transcript be like oh yeah actually that does fit with my memory actually has and it's an interesting experience to actually go and fly with one of these things because know there's a lot of your brain itself is also a probabilistic system and there's sort of ways you know The you know it could have been this way point it could have been this other way point it sounded generally like this in your head And so when you actually get that sort of suggestion of actually that's breezy it's oh yeah it was breezy And so actually how did they spell it Like how do you spell it What's the one around here yeah it's it's B-R-Z-Z-Y as opposed to bbr you know I don't know you know bbr EZY or something like that And so and when you go dial it into your you know your FMS then you'll actually see that it's right here and it's not someone at the other side of the country

Luka 24:29

do you find this easier to use than the, replay last a TC transmission option in glass cockpits?

A List 24:35

I think so because actually It's accessible for a much longer period of time And it's also now building some sort of shared understanding of what's going on of your flight So essentially as you continue to do your flight it's Okay let's say you know you should be writing things down in paper on your knee board But you know sometimes you call up and then they give you a clearance you're not ready for then immediately somebody else speaks on the frequency and you're just like okay I wrote down some of that clearance I'm not gonna enter it right now I'm like I gotta go lean my mixture or something like that And so you know and then in a minute or two you'll go dial it in and you'll of course you'll have read it back to the controller beforehand but you know do you actually spell those things And like you know where is you know where is you know the avenue or Avalon or VOR you know what is the actual like I said the right words back but like you know I don't actually know what the you know what those letters are I'm gonna go and just ask and basically that's gonna come up and be instantly available the just having an extra pair of ears actually does you know and this is based on this is based on flight experience you know of course I had some confirmation bias Hey I think it's gonna be cool but you know we when flew with it and actually it's pretty useful for a lot of these situations and it's made more useful But I think these this hearkens to some of the open questions which is you know Do you actually want it to hear a TC Do you want to talk to it directly know I think we're entering a Or do you also want it to talk to other pilots in the cockpit or just you and what know do you want it to be associated more So would people in the future want us to be associated more so with an aircraft to understand sort of the maintenance history of the vehicle for instance and be able to ask questions about the maintenance history and relay that to the pilot would it be associated more so with a person You know we're entering a an era where a lot of people are working with AI systems on everyday basis that start to understand more and more their individual life So you know there's a to be there's a likelihood that people are gonna want it more of an individualized type of communication experience But I You know we're going really carefully into this I think that's the I think that's the main thing I'd love you to take away from this question which is you know when there's an issue we go and add that to our set of scenario based testing that we have And whenever there's an update to one of these models we make a change to something related to the kind of the core structure We go and back test it with everything we ever hit When customer says Hey this happened and it was weird Can you go look at it Then we go and add it to the sort of the that set you know by because we're actually starting with flight training and this is actually a kind of a newer part of our product is this inflight kind of Beta testing And by the way we vet the people before we let them go fly with this because we know that there's a variety of types of pilots also in the world

What Problems Would GA Pilots Want Solved With AI

Jim 28:17

Alex, if we, you're in New York City right now. Let's say we got 10 GA pilots who fly in your area They're on the line right now, let's say, and we said, it looks like AI has the ability to help in a couple of different ways. if we got a whiteboard and we'd said, we want you to write down the 10 things where you are having a problem today that maybe AI could play a role in helping to resolve, whether it be in flight training or in real time flying, what would be the top five that you think there would be some agreement on where they may not know how the AI may work, but they're saying these are either complicated. Problems that I have to deal with, there may, it may relate to safety of flight, or it may be I can fly faster somewhere or more reliably. again, regardless of how it was done, what would these GA pilots generally agree on?

A List 29:18

I think the top two and this is funny cause this is reminding me of the wing episode where it's like we thought people would want medical supplies in the bush but it turns out they just wanted delivered coffee overnight you know within five minutes or something like that So the top two things would be from the entire internet being able to go and distill that and provide that to you while you're flying So with references so ways of being able to check that

Jim 29:43

for what? Re again, these are GA pilots. What's that value to them?

A List 29:48

Yeah So I'm getting indications My alternator failed you know what could have caused this one And what have do other people see this in this particular aircraft Is it really that risky You know I hate to say this I've been on a flight where alternator failed and I found out only later that the mags had been swapped One of the pair of mags had been swapped to E mags And I was like oh Didn't realize this was like a discontinue immediately thing until after I you know got to my destination things generally don't fail all at the same time So there's you know there's that but like you know it was a instead of being if I had that information I might have made a different choice

Jim 30:25

So real time maintenance. diagnosis and correction. you'd think that's low hanging fruit. is that something that I would assume those 10 GA pilots would say, gosh, if I have a problem and I get an instant recommendation of what to do about it, would that be top five?

A List 30:41

So that would probably be that would probably be number one The the you have to always take it as a grain of salt because this information that's found is often from a variety of sources So some of it's from Reddit Like the references will be here's what these guys are saying on Reddit And then some of it is like from actual pos that are found And actually some of it's from your POH right And and so like you know for your POH you hit a link you hit a you hit a button and then here's the page for your POH right then for those other things it's if you really wanna go and read Reddit you can hit the button and then go read Reddit right But that's like you know that's totally on you

Jim 31:19

that's a problem that they would identify with. They would say, I, I gotta figure it out pretty quick. I try, whatever I have in the flight deck right now. But boy, if I could get a better diagnosis and a better recommendation of what to do about it, and then we'll worry about how that could be better refined and more exact and more at another time. So what's another two or three problems that GA pilots in New York? Gary would say I could use some help on.

Peter 31:45

This technology can give pilots local knowledge and familiarity in an area where they're not accustomed to flying so that they can fly the way the pilots that are based at those airports already know how to fly. So, when a TC pronounces a waypoint, your system could just listen for a week on every a TC channel in every different region, and under understand how do the controllers pronounce the waypoints? And we're not guessing with phonics, we're definitively mapping those callouts to the actual waypoint. Or when a TC gives you A-A-V-F-R, visual waypoint as part of the traffic pattern or something. a new pilot to the area has to say, I'm unfamiliar, and then they have to get, a bunch of handholding from the controller in order to figure out where they need to go. This system could help them oh, what did the controller just say? What was that way point? What does it look like? where should I expect to find it? yes, you can with more workload staring at the sectional, you can piece those things together, but it's not nearly as easy as if you're already familiar with the area. And I could see a system like this helping with local familiarity of things like that, including noise abatement, procedures, and all of the ins and outs, all the fine see around an airport about no touch and goes after sunset and all these other things. It's not about making you a better pilot according to the regs. It's about making you a pilot who is proficient in that local area and as familiar with how things work there as any other pilot there. that's intuitively,

Jim 33:20

So having the familiarity of flying a thousand times to an airport versus, I'm about to fly to an area where I've never flown it. What are the net benefits of having that capability, of having flown it a thousand times? So one, you mentioned noise abatement, so I know the sensitivity we were talking about HPN earlier. Come on, flying there. You gotta be really mindful of the noise abatement policies of HPN. what are the other results that come from that familiarity?

From Language To Autonomy

A List 33:48

So you've got basically those were the experiences I also had in Oakland recently You know here what are the way what are the visual way points for Oakland Also flying down to Santa Monica basically Hey you know when does this airport actually close I'm sure there's like a time limit and they're like actually it's just about engine starts and takeoffs Don't worry about it for landings right the implications are that actually points of interest in PI reps are gonna become a lot more accessible You know imagine like you look at for flight right now and you have a section That gives you notes on a you know from individual pilots and remarks on particular FBOs and things like that Imagine that on steroids Being able to say okay why do you go here I go here for the burger and et cetera et cetera So you'd be able to bring in that knowledge and it's Hey be aware that the credit card you know the credit card terminals actually not working on this fuel farm right now So you have to go during the day and actually pay So you know the folks that I talk to in air traffic control are also you know they say Hey something like this could actually help with being able to get more PI reps Being able to spread that knowledge around and I think you know there those are some of the second order benefits The immediate sort of n is equal to one benefits You're the only person using this Also include things like doing checking your mathematics I think a lot of pilots would enjoy that cause many of them have made mistakes related to how do you actually enter a pattern or you know just remembering things Talking things through know out loud with a little bit of a little bit of a backup when things go wrong I think that is a major thing I mean you see in aviation you know have the adages like there's old pilots and there's bold pilots and no old bold pilots there are situations where you know talking through a particular circumstance and figuring out those outs are actually useful for figuring out how to avoid the the worst scenarios And I think one other aspect to add in terms of what Peter said is flying into different jurisdictions also includes things like machine translation And so now like you know of course people can do phraseology and so you get the IKO phraseology when you fly up to Montreal but you're not actually hearing what the other aircraft are doing and you're not actually you know gonna learn as much about the local procedures or how to get progressive directions to your particular FBO necessarily or what the procedure is in terms of how do I get gas here or things like that from your friendly ground controller versus you know you're gonna be able to fly from one country to another and go out to Frankfurt and rent an end number aircraft Go fly around Europe and be able to actually visit a bunch of different countries and be able to actually handle it a lot better Of course you're gonna be using the A KO phraseology for being clear to land You're not gonna be translating that and pretending to put on a an American accent for your for your for your French controller there But the but all the other aspects are things that can actually be a lot better supported through these sorts of systems And so there's a lot of possibility here I don't think we even fully know yet what people are gonna be using this I mean we continue to add things to the system Like okay can I send can I now send emails and messages and have this Sort of like you know as being a way that a way that we get information into the pilot while they're flying So imagine having you know there's some aerial operations that you could be doing from a commercial perspective Hey can you go actually go check out this part of the farm Or this is going on with the this here's a update to your routing Can you go check out this other aspect of the power lines for instance And you can receive that information either through a handheld radio or process through text And so you know there's a variety of things that we're bringing to the platform So you know I think we're at the very earliest phases of this It's the it's only begun to be possible now through the variety of connectivity In fact we figured out only because of the compute that's now available on an iPad for doing automatic speech recognition And then on the ground it's only starting to become helpful Now eventually you know forward five years you're gonna have a system that is probably and so lemme just break it down Currently we break it into two different models There's a model that turns something into text from speech and so that takes all the a TC information and then turns it into text and then you take that text and you send it to the ground and then does all this magical processing And sends it back up to the plane and then it speaks it So you know there's a future where actually what you're doing is keeping it as audio And when you keep it as audio and then you actually turn it into a turn it into a mathematical representation that's not text per se but it rep if you were to you can turn it back into the audio That sort of deals compression algorithm when you are able to do this you get a lot more performance and you might be able to have a very kind of engaged conversation in the cockpit that's able to understand a lot more So there's this is only the very beginning and I think the performance is going to increase I think the real question is like how do you provably build out a system that works really well for pilots and also gives reassurance that it's that it works in an excellent way

Luka 38:43

and I would also add, how do you make the leap between, treating aviation and flying as a language problem by, turning a TC into words and analyzing words and turning that into a situational awareness into something that understands, geography, that understands the relative position of all the prioritized aircraft around ownership. and with respect to terrain and airspace and all of these things. is that a natural evolution of language models? Is that a, an interface or an architecture problem? Or how do we make that leap?

SkyGrid sponsor segment

A List 39:18

I think so I mean it's now a good time to I mean I would love to talk to you a little bit about a vision for how you could actually get automated flight rule traffic in the na in a way that's compatible with the existing system So I think one of the ways that you could actually do that is by having kind of drop in a communication systems that are able to augment and scale the workloads of pilots and air traffic controllers You go and provide a voice-based interface that's totally reverse compatible with the existing voice-based interface And then actually you go in and you start to automate out a lot of the work one of the things that's really interesting to me especially listening to your podcast that flight rules use space and the like are being designed for the future workloads which are really important and are going to be large it's a process of building alignment and getting all the nerds together in the room to make sure that when you have this amount of traffic that you're not gonna have problems I think an a another way of looking at this in terms of a compatible with the existing of mixed use airspace is that you have communication systems that are able to go And either communicate via voice or also communicate via text What's really interesting about the way that the existing language models work is that when they think that something through you're actually getting a textural representation of how they thought about it And that's very different than a mathematical representation where for instance with acas you might have a probabilistic environment where maybe one of those lobes is the most optimal for the increase of the separation between the aircraft And so that therefore you go there versus thinking some of these things through But fundamentally the air traffic control system so much of it has to do with communicating about current states of things So for instance if you're going into land at an airport And you know maybe there's some debris on the runway want there to be a way for you to say okay everybody hold off and you know go do you know you do right 360 you do right 360 And you wanna be able to have that be a thing that can be spoken by the existing controller and ultimately through some sort of supplementation and augmentation maybe then yeah you'll have a an AI system that once you know that the controller can improve this is the response Okay Yeah There's debris here Yeah Just hold everybody and then it'll go and actually hold people But I think this is something that gets built out really gradually just like people shouldn't go and you know Totally trust that the waypoint that the AI heard or go receive the clearance from the a TC for you not listen cause you're gonna listen to your favorite podcast instead And and then go listen to it and then put it that directly into your you know into your PFD or your MFD And you know instead actually I think people build out this trust slowly just like you the use of AI tools is actually scaling out the efficiency and the controllers and the remote pilots you know gradually And I think this sort of builds a combination where ultimately you can start to use digital interfaces in order to be able to transmit that same information on the back channel and have some of those things be automated So the automated flight rules concept is excellent You have different interactions with different vehicle classes So for instance if you have VFR and automated flack rules you probably need to give a big for somebody doing their you know their city tour or the you know spontaneously doing a left 360 over 10,000 feet over Chicago And you are you're interacting with IFR traffic probably a little bit less separation if you're interacting with other automated flight rules traffic really close separation maybe right behind somebody in behind them in their wake But the way that you get to being able to support that in extreme mixed scenarios and extreme mixed airspace is to be able to have voice systems that are reverse compatible because they work today as you scale out remote pilots which is being done And then controllers which is an effort to be done currently.

Luka 43:20

Since we're talking about automated flight rules, now is a good time for a short segment with our sponsor Sky Grid. Here's a quick conversation between Jim and Brendan from Sky Grid, on how to scale air traffic management in advanced air mobility.

Jim 43:32

Brendan, how can we scale air traffic management to support advanced air mobility and autonomous aviation?

Brenden 43:39

Well ultimately scale comes from automating the routine and digitizing the rules that we have in place today. So, you know, today's system is tactical. Voice centric and very human limited. So in order to achieve high tempo, low altitude operations, that will quickly overwhelm what we have in place today. A lot more automation is really necessary throughout all phases of flight, both the strategic and tactical phase. So automating the flight planning, scheduling, coordination prior to departure, as well as the detection and resolution of conflicts in flight and ensuring separations. Ultimately, this will require the introduction of new airspace, constructs, and a new mode of operation that's really tailored to highly automated flights. So think automated flight rules or a FR that standardize both the intent sharing and the machine readable constraints. This pairs with the increased automation, onboard the aircraft with certified ground services for traffic awareness, constraint management, and deconfliction such that a TC and voice-based communication isn't the bottleneck because ultimately we won't be able to scale a a m operations by adding more radio calls, but rather by adding more automation to what we have today.

Jim 44:53

How do we develop advanced term mobility, digital infrastructure with the right level of assurance for safety critical applications?

Brenden 45:01

Well, firstly, we need to treat data and services like we do avionics. We need to certify them such that they can be used reliably. You know, this requires us to move away from those kind of best effort applications to Certified third party services with defined assurance level. So, you know, I think automated data service providers under part 1 46 as introduced by the FAA and verifiable integrity latency and availability of that data. So we need to architect for defense in depth, which you know, means authenticated data sources, deterministic pipelines, health monitoring, and you know, fail operational fallbacks for optimal security. As well as apply, you know, aviation great practices to software development. So requirement traceability, hazard analysis, verification and validation, and continuous assurance across releases. So at the end of the day, if it informs safety of flight decisions, which is a must, if you're going to rely on increasing levels of automation for scale, it needs the same rigor as a flight computer does today.

Luka 46:06

And now back to our conversation with Alex. When it comes to connectivity, do you see the future of, cockpit AI as starlink plus some, cloud-based brain? Or do you see that as an iPad with some local intelligence? what are the pros and cons for the two models as it relates to things that you can do with the product and, certification, et cetera?

A List 46:32

Yeah I really do see it and this is where I have two kind of divergent philosophies here So I see it as a pilot assistance being done using ground based intelligence and using the frontier level of intelligence that is really excellent there The real kind of hack is that you don't need a lot of bandwidth in order to be able to make that connection happen And so you know on a 56 kilobit modem you can transmit text just fine You can send that fax no problem if you're actually trying to the audio that's where you start to have the hiccups the mixed type of architectures that are gonna be possible for this sort of application but also for other sorts of remote outback control applications are ones where actually that intelligence matters a lot this does actually differ than my opinion for ultimately you want to be moving around super valuable merchandise or people you probably want to have that level of intelligence on the vehicle this is why I believe that this is a little bit more of a long-term sort of vision I think eventually you're gonna end up having vehicles that are able to communicate independently you know based on where this technology is going I think the that level of support because in the former the human is the primary and final authority for the operation of that flight And in the latter that flight needs to run independently of any individuals So yeah that's my view on that

Luka 48:02

it matters a lot because it really determines what company you're building, ultimately. And what kind of, worldview do you believe in is aviation connectivity, quote, good enough. to treat cockpit AI as cloud-based connected software or, or do you take the stance that serious cockpit systems must assume degraded or zero connectivity. and, that obviously has ripple effects on what the scope of the product is, but trust as well. And so it is a delicate balance, no doubt.

A List 48:35

And there's some value you can provide without that level of intelligence So for a lot of strategic questions you can wait 30 seconds You know I go back and watch our recordings from our flights You know the in route phase it's a three and a half hour flight The in route phase you know there's massive gaps between people talking to each other It's like you're fishing You're just hanging out You got time you're spending time The kind of the approach in terminal phase that's where things really speed up You know in an in route phase for a general aviation pilot fly in a Cessna there's a lot of times where they have one kind of sketchy bar of connectivity there's actually technologies these days that work really well for that HTTP three is able to resume connections really transparently as you're swapping from cell tower to cell tower you've also got the ability to go and buy a starlink throw it on your windshield Now you've got a little bit better connectivity It's pretty excellent although in a steep turn you might lose it And obviously there's competition coming to that space with Amazon Leo and things like that So you know I think connectivity starts to become really good And you know even with a small amount there's a lot of value you can provide in the copy run now through these intelligence systems

Jim 49:43

How is, AI being used today with GA, commercial or military? What are some basic uses of AI either in the training or in the flight deck that you can tell us about Alex?

A List 49:55

So I think everybody starting to wake up that there's a lot of language models being used in targeting and in the procedures for identification of targets and the utilization of surveillance information like a obviously an act of conflicts where there's way more information than there is to be processed in terms of the actual information that's being fed through language models in flight there's say an effort to go and if you're gonna do a TC modernization you wanna make sure that you know What are the issues people having now then you wanna be able to quantify whether or not you've made the system better or not an effort related to going and capturing using automated speech recognition going and capturing you know all the instances of various problems that came up so that after you roll out a modernization to the air traffic control system can you actually demonstrate that you have better performance than you had before So actually that's being done Using using you know speech to text and some level of language models there's also you know there's also of course just a number of of civilian based apps that are using AI in order to analyze flight track information and you know there's you know give you a push notification when something's going on you know some like hot pursuit on a tar tarmac or something like that Then there's you know of course what we're also doing which is related to training which is our primary effort And so we actually go and provide a system of building out curriculum and then delivering that we can go and automatically generate You know curriculum to a particular spec one of the things that's interesting is that we didn't expect this but we actually found that instructors are know we thought hey we'd build a curriculum then people would go and use It's actually instructors want to use this system to go generate their own curriculum and then provide it to their students And so that's another sort of avenue for all of this And the value is that you can go and tailor day not just a group of students but also individual students and allow that individual student to have a lot better of a kind of a of an experience going and using their flight simulator to the maximum physics extent to actually learn really important elements

Jim 52:08

Alex, you have a very cool past. you're a serial entrepreneur for sure, and you've worked at Boeing, you've worked at Merlin, you've worked at Beta. drew you to this space? what was the draw and what excited you?

A List 52:23

So you know fundamentally I am a builder and I think there's a you know you don't see platform shifts all that often And so you know seeing a platform shift you know Made me and I think this would probably be different than probably most of the people on your podcast right But like you know seeing that there's a platform shift Hey language models are starting to be a thing Can we apply this to aviation a friend of mine had a had an aviation accident and I wanted to go potentially go and build something And so you know took what I had saved up to buy an airplane instead started a company and to try to make flight a bit sa safer And so went into this with that in mind you know with otherwise with an open mind And so I've been taking that approach ever since So the kind of combination of a new technology in a space that I am really passionate about and I

Jim 53:15

by a friend who had an accident where

A List 53:17

that's right

Jim 53:17

there's probably a better way here.

A List 53:19

be a better way and hey maybe maybe there's some sort of way to improve aviation safety to some degree And so you know we take we I take all that forward into try to build something that is is meaningful for people Obviously it's a process by which you need to talk to customers and figure out what's needed the you know the time of being able to maybe just start something because you've got a platform shift that window might be closing for people a little bit But hey as somebody who has been building their entire life it's a really it's a really great it's been you know it's a really great journey to be on

Business Model And Wrap Up

Peter 53:51

the journey that you've been on, it's interesting to start with aviation safety as the initial impetus that sort of created the mission for you. but then as you delve into it more deeply and you pick up this very powerful new technology and try and figure out, what does the product look like that this technology can now enable? it really unfolds that there are a lot more facets to this, this overall goal of aviation safety than, simply being better at memorizing checklists or, doing the very core piloting, operations, the things that are defined in the regs, but that there are, in addition, these, intangible, aspects that contribute to it, that are around, familiarity and around reducing workload. and around things that are not easily acquired by pilots today. without them actually flying into that scenario or flying into that airport or that region and just experiencing it firsthand. And then the second and third time they fly through it, the workload is much lower and the safety goes up. but that, that really identifies a big opportunity, I think, and a way to solve it in a way that, that really wakes people up and, is a call to action for them to say, you know what? I actually am going to, change my way of doing things. I'm gonna try adopting this.

Luka 55:11

Speaking of which, Alex, how do you then build a business out of this technology in general? Aviation. I.

A List 55:17

sorry. You know, I'm, I am, I'm lucky to be meeting with you all. There's a, you know, I'm, I am curious what you all see in this space, but I think one of the challenges in building in general aviation, or building an aviation in general is that you look at it and you, from the beginning, you're like, oh, this is looking pretty good. You've got, aviation is 4% of the world economy. You know, that's pretty good. and then when you start to look at it, you're like, okay. And then you've got 200,000 uh private pilots, 500,000, like non-student pilots. Okay, what are those numbers? It's oh God, two, two times 10 to the second. Okay. That two times, 10 to the, 50. You've got alright. Then you're like, okay, what is the, what's, what's the number you get outta here? You're like, okay, this is 10 to the, this is 10 to the seventh. You're like, okay, we got like, Tens of millions of dollars. You know, you're addressable and you're like, oh, bummer. You know, it turns out, you know, a venture capitalist, they need to hear, you know, they need to hear 10 to the eighth. And so, okay, let's go back, you know, let's go back to the drawing board here. Okay. You start to look around and you're like, okay, what else is there? Here? And you're, and then you look at the pilot statistics and you say, okay, actually there's, turns out there's like already 450,000 remote pilots, and I think this is probably what you guys also see sometimes. And so you're like, there's 450,000 remote pilots, and that's growing by 60,000 a year. And a positive second derivative. You're like, this is actually gonna be, you know, a pretty big industry. You're like, and then you're, and then you're like, already that's, hey, that's the 10th to the eighth. And then you're like, already, then you layer in, you know, counter UAS and things like that. and that are in markets that are very underserved. And now you're like, Hey, there's, there could be a really, there could be like a real business here related to training and operations. But anyway, you folks see the, you see the thousands of decks a year. I am really curious, like what business models make sense to you all?

Jim 56:58

So one of the things we do in our company is we do something called a challenge briefing. So what we actually do is we put experts in a room. Let's say you are in the briefing right now, and there's no question there's full 450,000, potential customers, but what are they willing to pay for? and safety historically has been something where it's tough to make money on safety. and I'm particularly enamored by Peter's comments of, you can create every potential airport that I'm gonna fly to, you have the experience of flying it to a thousand times, what's the value to that pilot? What's that pilot willing to pay for that AI is going to enable?

A List 57:38

so what's the value of a chat GBT subscription or an philanthropic cloud subscription? It's like you have access to the world's information. You have a sort of a kind of a thinking partner. And actually that's what's really interesting is that we've taken a lot of the value there and tailored it to a really specific application, which is, you know, aviation. We've put a lot of work into making sure that that information is delivered in a way that is reliable and also suitable to the context. You know, there's a lot of work that goes into synchronizing voice in and out of a TC and interruptions in ways that kind of work within a cockpit in context and environment. So there's that aspect. There's the training aspect, which is starting to come together in terms of what it's gonna look like. And a lot of that looks like, hey, there's a lot of organizations that actually do training they want to, and they want to go and provide that training to other people. It's interesting that the types of folks that come out are really, varied. So we've got a partnership with sporty, we've been doing some visual training for people, and then we're also developing some instrument scenarios together that's gonna be available from them on Shirley. We've got a, there's a set of, it's a variety, right? Because you also have people who are, you know, enthusiasts for aircraft and want to teach people their way to go and actually start up an aircraft that they might find in DCS or something. And then there are, and they offer that as a curriculum. And then there's like, you know, there's a lot of, flight instructor and flight instructor programs, flight instructors who are associated with programs that have quite a number of students and they wanna build curriculum. And so that's one aspect of training that's starting to work out. There's starting to look at, the UAS space too, because we can go and deliver to service members that might be in a barracks. The tactics and the strategies being used in, in warfare changing really quickly, you know, specifically because of the Russia, Ukraine conflict and the, these are things that people actually need to become, you know, the skills atrophy. But we can actually go and deliver the modern tactics through a kind of an a, an iPad that somebody has already in the barracks there. So, you know, I think it's, you know, what's really interesting is that we're, we're, we are, you know, we're figuring out just like a, as a startup goes, but it's, I think a really interesting time to be doing this and.

Jim 59:58

if you're making, let's say 30 million a year in a couple years. Where is 80% of your revenue company from, do you think

A List 1:00:05

That's a,

Jim 1:00:05

companies like yours?

A List 1:00:07

yeah, no, that's a good one. So, you know, I think, you know, I think there's a really good opportunity, there is a really good opportunity in flight.'cause that's one product that you scale out. The actually training delivery is another product that gets scaled out. You know, I think a lot of that's gonna come from progress. You know, increasingly that's gonna be coming from folks working in, I would say in the UAS space. You know, right now people really love flying their sims for instrument training if they're gonna be doing something there. And so instruments are a really big one. But I also think that it's just because everybody wants to be able to deliver some sort of content. It's not about, not necessarily about how, we deliver in our particular curriculum. It's about giving kind of a creative tool for somebody else. You know, I'd never thought I'd actually be saying this because I used to think of it, as an anti-pattern. I'm gonna go build a platform for X, Y, Z. it really does seem like in this era where you can go and you, go and be either a feature. You know, if you wanna go try to be a feature in somebody else's platform, like let's say we say, Hey, we wanna build a speech system and go sell it to, I don't know, the avionics maker or one of the drone operators. Hey, you know, in the era of Claude, if it's easy to build your thing, then they can just go build it themself. You know, why not? They'll just go, actually just go auto generate some code. They have the resources. They'll put a bunch of people on eventually, like, you know your stuff, especially if you're involved. In a kind of a front office job, You know, the, you know, surely ends up being a front office type of, you know, like actor because of, you know, there are trades and like, you know, are behavioral emphases that kind of, know, that we put in order to make it, you know, in order to work stably within our system that end up being opinionated. And so, you know, that ends up being a thing that needs to be carried forward somewhat and can be adjusted in a lot of circumstances. it's something that would end up in somebody else's, you know, portfolio. If they're going in, let's say, offering Shirley inside of avionics. And so the way you build a business here in the, in, as far as I, you know, as far as I see it, is to have a moat related to data and also have a moat related to the actual operation or provision of a particular service. And so, you know, I think training is a really excellent one, especially if you can allow other people to deliver training through what you do. And that's starting to become really, more of a, more of an emphasis for what we do here and becoming more clear as the right way to do it. And then the other one is to provide a direct relationship of a customer through, for instance, like a tablet based experience that comes with all these other risks. Of course, startups are better positioned to be able to take these risks than big companies. Like, you know, it used to be Boeing with four flight, but not anymore. it's, it that continues to be the case, but all that for sure needs to be figured out.

Luka 1:02:52

Yeah, I would add to things that are less attractive from a business model, perspective as it relates to, building a business in ga. if you're starting with a tiny enthusiast user base, that's, that's not ideal. Also, if you're immediately going after some safety critical dependence, or. Need certification gates right off the bat, or you require some, bespoke heavy integration with the OEMs or avionics companies or just broadly you have some OEM dependence and you don't really have the leverage. those are pretty bad starting points.

A List 1:03:29

Right. And so what else do you guys see as being, you know, I was impressed of, you know, for instance, the ED from Drone Hand came on to your show and he was talking about how actually his wedge is actually being an app on the DJI controller and he is able to bring that machine learning to that particular app in that environment. He's able to scale this out to all these different farmers. That seems like a really excellent kind of approach to me. I'm wondering what other sort of approaches generally that you guys have been, you know, impressed with you since after all you probably have pattern matching than almost anybody else on the planet related aviation startup businesses.

Luka 1:04:04

I appreciate the comment, you wanna see a clear operational pain that people are addressing, and a, recurring software like revenue opportunity, not requiring certification or SDCs. you come in, with a narrow wedge that then has the potential to expand over time. You have some kind of a data advantage that compounds with use. yeah, those are the things that I would flag. and also avoiding the failure mode of becoming a, a services company in disguise. where you start with a genuinely interesting technology, but because of the market fragmentation and, slow rate of speed and regulation and the conservative buyer base, you end up being pulled into, different bespoke type of work and. Your revenue does come in, but it's all services heavy. your gross margins are, nothing to write home about, and your scaling is constrained by labor, right? Those are the things that, as a VC are flags, red flags.

A List 1:05:10

Yeah, that's really interesting.

Jim 1:05:11

I do wanna say my gut reaction right now, and I've partly phrased it with the questions I'm asking, what are the problems you're solving now? what's the user group you're targeting? What are the problems they have? My sense is they're pretty meaningful problems, both on the efficiency side, the safety side, and then list that out. And then my guess is there's low hanging fruit there. What's the low hanging fruit that you could address? And ideally you're doing it through some kind of a, software solution, versus services based as Luka said. But I'd start with that. if you can address that, you can articulate that you get a lot of head nodding from GA pilots in that area. And then ideally find comps, find other ways that they've used technology then you've hooked. potential investor with big problem been done before, ideally subscription based. Now you got my attention and I respect that you're leaning into something here and you're trying to create a new way, and the first guy through the wire is the bloodiest, but I respect that. you're trying to solve what seems to be some meaningful problems.

A List 1:06:17

I know this isn't Shark Tank here, so, you know, we can keep it, we can keep it like, you know, on topic. I really appreciate that, Jim. I think like, you know, I am very curious. I mean, you, I think from your perspectives you would be sitting at a place where you see everything and, you know, just having that point of view to be able to bring something I think is just, you know, really cool to be able to bring that into the podcast, you know, on occasion to be able to bring in your particular insights because you have a lot of knowledge based on what you see. So I think, you know, wherever you could bring that in, I would encourage that because you just have that, you know, that expertise.

Jim 1:06:55

serial entrepreneur? Very technical guy. been with some very cool companies. started cool companies. what message would you wanna send out to our entrepreneurial audience

A List 1:07:07

I would say that you have to fundamentally build something people want. You need to take time and, you know, talk to your customers, talk to potential customers. A conversation with a customer is never a waste of time. You know, go have that conversation. I try to move most of my kind of, you know, non-customer related interactions a very segmented part of my week. The rest of it is interacting with customers and building. And so, so that's really my approach to it on how we do this. I think like you're, one way or another like, just go talk to people. you will go try to figure it out. So it's a, it's a journey.

Jim 1:07:47

Good for you. You're leaning into it, man. We respect it. thanks for joining us, Alex.

A List 1:07:52

Thank

Luka 1:07:52

that we did?

A List 1:07:53

you, all three of you. It's a real pleasure to join you here and, look forward to continuing, seeing your pods and talking to you folks going forward.