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

In this episode of The Vertical Space, Alex List — CEO and founder of FlyShirley — walks through what AI in the cockpit actually looks like today. Not the vision, but the reality: a ground-based language model accessed via iPad that helps pilots handle strategic, non-time-critical tasks like looking up service bulletins mid-flight, transcribing ATC clearances, finding alternates, and synthesizing information that would otherwise require digging through a POH while managing weather and workload.

Alex is candid about where the technology still falls short and articulates a clear architectural thesis: frontier intelligence lives on the ground, state management lives on the device, and a 56-kilobit connection is all you need in between. The conversation covers pilot dependency risk, the liability problem for advisory AI systems, and the hard math of building a defensible business in a market that is, honestly, pretty small.

Key Topics

• What AI in the cockpit can reliably do today — and what it still gets wrong
• Why ground-based intelligence beats onboard compute for pilot assistance at current technology levels
• How to design AI copilot systems that build stronger pilots rather than creating dependency
• The liability problem for aviation advisory AI — and why it remains unsolved industry-wide
• ATC transcription in practice — how AI handles noisy frequencies and ambiguous clearances
• Why giving a pilot arriving at an unfamiliar airport the experience of having flown there a thousand times is one of the most valuable AI applications in GA
• The business model math for GA — why remote pilots and UAS training may be the real opportunity
• How automated flight rules could be introduced as a drop-in compatible layer within the existing ATC system

+ Read Full Transcript

Alex List: 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. Training is a really excellent one, especially if you can allow other people to deliver training through what you do. And then the other one is to provide a direct relationship with a customer through, for instance, 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: Hey everyone, welcome back to The Vertical Space and our conversation with Alex List, CEO and Founder of FlyShirley — a startup developing an AI co-pilot to assist pilots in the cockpit. Alex describes Shirley as 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. One particular compelling idea: AI could give pilots arriving at an unfamiliar airport the equivalent experience of having flown there a thousand times before.

Luka: Alex, welcome to The Vertical Space. What is it that few in the industry agree with you on?

Alex List: 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 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. ATC is about communication and less so about coordination. Terminal is a really messy airspace. Rather than having a controller potentially need to have another keyboard to communicate with automated flight rules traffic, we believe you can actually automate ATC 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: What actually is the state of the art when it comes to AI being used in the cockpit?

Alex List: Language models are improvisation machines — give them a script and they're able to fill in various roles. One of those roles can be a co-pilot. These models aren't necessarily that much smarter than individual people but they think a lot faster, never get tired, and never get affected by turbulence. They have crazy long-term memory. What currently works with pilot assistance technologies are things that are easy to check but hard to look up — strategic and less tactical things. There's an example where we're flying from Sacramento back to San Jose and getting shadowing on our ADS-B. We asked Shirley what's going on and it was able to find a service bulletin related to it. It said one thing that was correct and one thing that wasn't correct — it highlights how there can be one piece of valuable information and another piece that's not useful. The pilot in the seat synthesizes it for strategic and less tactical things. In the training space, these systems work extremely well related to storytelling — you can put somebody in a scenario and have language models provide storytelling. You structure that with decision analysis and put a large amount of structure around it and you've got something really useful.

Training and Low Hanging Fruit

Jim: Do you see the application of AI as a copilot being more viable right now in the simulation world versus real-time flying, and what's the low-hanging fruit?

Alex List: In the training space you're able to give instructors superpowers — generating adaptive training for individual students, recording a flight and having it debriefed and ready for the instructor to view. Most people don't actually graduate with the minimum number of hours required. There's an excellent opportunity to have people prepared for their lessons. In flight, our experience flying around the Midwest and California has shown how useful these things can be related to finding alternates and figuring out plans. When you're in a single-engine piston aircraft trying to figure out what you can actually accomplish on a particular flight, you ask the AI "what are other people doing in this situation?" It goes finds out on Reddit how do people normally cut into the Central Valley. When problems come up and you're in the soup and getting bumped around by turbulence, it's not the right time to be going through an operating handbook to page 274. The co-pilot just goes and finds it for you.

Luka: What are some concrete tasks that AI in the cockpit can reliably handle today?

Alex List: A year ago you might have had problems related to hallucinating. Those have disappeared. The tooling has also improved — we give it a research tool that can scrape the internet and come back with information, even watch YouTube videos very quickly. What you actually want for an in-flight usable copilot is that you want state management to be done on the device — your iPad — and you want the thinking to be happening on the ground. You're basically able to continuously supplement with tools and memory, pull out things that happened in previous flights, set reminders, ask really deep questions, do mathematical computation to determine times of descent. You've got something that feels like a pretty intelligent entity that can think a lot quicker in a lot more circumstances than somebody holding onto a yoke dealing with weather.

Building a Stronger Pilot vs. Pilot Atrophy

Peter: What is your philosophical approach to product design such that it builds a stronger pilot rather than creating pilot atrophy and dependence on the AI?

Alex List: There are a lot of parallels to GPS — the rollout involved people forgetting how to do navigation using compasses and landmarks because now you're just following a magenta line. This is a tool for cognitive offloading. At the end of a long flight where you're having trouble subtracting 180 from 230, there's an additional layer of support in the cockpit. This gets into questions around training, automation management, and dependency that people really need to look at seriously. We are very serious about making sure that the things the system can do are done, and those it's not equipped to do — we just don't provide information on those things. We're being quite cognizant in how we roll this out. There are implications related to liability for systems like this. Pilots are used to using whatever information they have at their fingertips to make the best judgment call they can at the time. I fundamentally believe in pilots' ability to understand what the utility of the tool is.

The Hard Tasks

Luka: What are some of the tasks that are still hard for Shirley, or for AI in general? How good is the technology at interpreting messy ATC exchanges or reasoning spatially about what it hears on the VHF radio?

Alex List: There's evidence that for language models, if you have a lousier image but a smarter model, that's actually better. There's a similar effect with speech understanding. We're starting to have the latest frontier models — and that kind of combination allows you to have good comprehension. The model seems to know when information coming to it is over-compressed and is able to ask follow-up questions. That aspect is increasingly well solved. The harder problems are that models are fundamentally disembodied. We're flying over Indiana and ask "what's this oil refinery off my right?" — the model is going to pick the one that is culturally most salient unless it has a tool to search for points of interest. It can do cosines and signs all day but it's not going to know what "off your right" means or "at your three o'clock" currently. Taxi chart understanding — you'd have a hard time having these models trace a corn maze. We took the best off-the-shelf open source model with the best compute we could fit in an airplane and it still wasn't as good as having an automatic speech model that we ran on an iPad using compute done on the ground.

Guardrails and Liability

Luka: What's the case against it? What's the chance the AI hallucinates the wrong clearance or invents what ATC probably meant?

Alex List: Those are absolutely the right questions. What's really interesting is — if you look at design assurance levels in systems that exist today, the most non-deterministic element in the air traffic control system is the one that's viewed as infallible. On the other hand, humans have skin in the game — pilots are flying in the airplane and make the ultimate decision. The liability for advisory systems is totally unsolved industry-wide. It's similar to the General Aviation Revitalization Act of 1994 — Cessna was getting sued into oblivion. At the time, a hundred thousand dollars of the sale price went to liability insurance. For now, this AI system is an information layer — definitely not taking action in the cockpit, but soon proposing flight plan changes to accept. That's the same approach for rolling out drop-ins for air traffic control: start with summarization and transcription. The human takes a look at the transcript, does this actually make sense given the context, and then the system says here's what I recommend in terms of the course of action.

ATC Transcripts in Practice

Luka: Specifically in the case of transcribing ATC, it's awfully easy to miss one letter or number. How does the AI handle that?

Alex List: The most important aspect is to use this as a supplemental system for things that are easy to check but hard to look up. In the case of what the controller said — you're like, what was that? Then when you scroll back in the transcript you might see it does fit with your memory. Your brain is also a probabilistic system. It could have been this waypoint, it could have been that waypoint — it sounded generally like this. When you actually get that sort of suggestion "actually that's BREEZY" — oh yeah, it was BREEZY. How do you spell it? B-R-Z-Z-Y. When you go dial it into your FMS you'll actually see it's right here and not somewhere on the other side of the country. It's an interesting experience to fly with one of these things. Just having an extra pair of ears actually does help. When there's an issue, we add it to our set of scenario-based testing. Whenever there's an update to one of these models or we make a change, we back-test it with everything we've ever hit. We vet the people before we let them go fly with this because we know there are a variety of types of pilots in the world.

What Problems Would GA Pilots Want Solved with AI

Jim: If we got 10 GA pilots who fly in New York on the line and asked what problems AI could help them with, what would be the top ones?

Alex List: The top two things would be: from the entire internet, being able to distill that and provide it to you while you're flying with references. So real-time maintenance diagnosis and correction — that would probably be number one. My alternator failed, what could have caused this? What have other people seen on this particular aircraft? Is it really that risky? I've been on a flight where an alternator failed and I found out only later that the mags had been swapped to E-mags. If I had that information I might have made a different choice. Things generally don't fail all at the same time, but like — if I had had that information.

Peter: This technology can give pilots local knowledge and familiarity in an area where they're not accustomed to flying so they can fly the way pilots based at those airports already know how to fly. When ATC pronounces a waypoint, your system could listen for a week on every ATC channel in every different region and definitively map those callouts to the actual waypoint. A new pilot to the area has to say "I'm unfamiliar" and gets a bunch of handholding from the controller. This system could help: what did the controller just say? What was that waypoint? What does it look like? Where should I expect to find it? 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.

Jim: So having the familiarity of flying a thousand times to an airport versus flying to an area where you've never flown — what are the net benefits?

Alex List: You've got the visual waypoints, knowing when the airport actually closes, knowing noise abatement procedures. Points of interest and PIREPs become a lot more accessible — imagine ForeFlight's notes from individual pilots and remarks on particular FBOs on steroids. Being able to say: "Hey, be aware the credit card terminal isn't working on this fuel farm right now." Flying into different jurisdictions also includes machine translation — flying to Montreal you get ICAO phraseology, but you're not actually hearing what other aircraft are doing. You'll be able to fly from one country to another and actually handle it a lot better. There's a lot of possibility here I don't think we even fully know yet. We continue to add things to the system — can I now send emails and messages, having a way that information gets into the pilot while they're flying, receiving updates to routing. We're at the very earliest phases of this.

From Language to Autonomy

Luka: How do we make the leap from treating aviation as a language problem to something that understands geography and the relative position of aircraft?

Alex List: One of the ways you could get automated flight rule traffic in the NAS in a way that's compatible with the existing system is by having drop-in communication systems that augment and scale the workloads of pilots and air traffic controllers. You provide a voice-based interface that's totally reverse compatible with the existing voice-based interface. Then you start to automate out a lot of the work. Fundamentally the air traffic control system is so much about communicating about current states of things. For instance, if you're going in to land at an airport and there's debris on the runway — you want there to be a way for you to say "everybody hold off, do right 360s." That can be spoken by the existing controller and ultimately through augmentation maybe an AI system handles it. But people build out this trust slowly. The automated flight rules concept is excellent — you have different interactions with different vehicle classes. The way you get to supporting that in extreme mixed scenarios is to have voice systems that are reverse compatible because they work today, as you scale out remote pilots and then controllers.

Luka: SkyGrid sponsored Q&A on scaling air traffic management:

Brenden (SkyGrid): Scale comes from automating the routine and digitizing the rules that we have in place today. Today's system is tactical, voice-centric, and very human-limited. In order to achieve high-tempo, low-altitude operations, a lot more automation is necessary throughout all phases of flight — automating flight planning, scheduling, coordination prior to departure, as well as detection and resolution of conflicts in flight. This requires the introduction of new airspace constructs and a new mode of operation tailored to highly automated flights — Automated Flight Rules or AFR — that standardize both intent sharing and machine-readable constraints. This pairs with increased automation onboard the aircraft with certified ground services for traffic awareness, constraint management, and deconfliction, such that ATC and voice-based communication isn't the bottleneck. We need to treat data and services like we do avionics — certify them such that they can be used reliably. Automated data service providers under Part 146, verifiable integrity, latency and availability. Architect for defense in depth — authenticated data sources, deterministic pipelines, health monitoring, and fail-operational fallbacks. If it informs safety of flight decisions, it needs the same rigor as a flight computer does today.

Luka: When it comes to connectivity, do you see the future of cockpit AI as Starlink plus some cloud-based brain, or as an iPad with local intelligence?

Alex List: I see pilot assistance being done using ground-based intelligence — the frontier level of intelligence that is really excellent there. The real hack is that you don't need a lot of bandwidth to make that connection happen. On a 56-kilobit modem you can transmit text just fine. That differs from my opinion for ultimately moving around super valuable merchandise or people — you probably want to have that level of intelligence on the vehicle itself. In the former, the human is the primary and final authority for the operation of that flight. In the latter, that flight needs to run independently of any individuals. For a lot of strategic questions you can wait 30 seconds. In the en-route phase of a three-and-a-half-hour flight, there are massive gaps between people talking. It's like fishing. You got time. The approach and terminal phase — that's where things really speed up. HTTP/3 is able to resume connections really transparently as you're swapping from cell tower to cell tower. You've also got the ability to buy a Starlink and throw it on your windshield.

Jim: How is AI being used today with GA, commercial, or military?

Alex List: Language models are being used in targeting and in the procedures for identification of targets and the utilization of surveillance information in active conflicts — way more information than there is to be processed. In flight, there's an effort to capture using automated speech recognition all the instances of various problems in ATC so that after you roll out a modernization to the ATC system you can demonstrate whether you have better performance than before. There are civilian-based apps using AI to analyze flight track information and give you a push notification when something's going on. And of course what we're doing — related to training, which is our primary effort. We provide a system of building out curriculum and then delivering it. We can automatically generate curriculum to a particular spec. Instructors want to use this system to go generate their own curriculum and then provide it to their students. You can go tailor not just a group of students but individual students and allow them to have a much better experience going through their flight simulator.

Jim: You're a serial entrepreneur who's worked at Boeing, Merlin, Beta. What drew you to this space?

Alex List: Fundamentally I am a builder. You don't see platform shifts all that often. Seeing that language models are starting to be a thing — can we apply this to aviation? A friend of mine had an aviation accident and I wanted to potentially build something. I took what I had saved up to buy an airplane and instead started a company to try to make flight a bit safer. The combination of a new technology in a space that I am really passionate about — that's what drives it. Obviously it's a process by which you need to talk to customers and figure out what's needed.

Business Model and Wrap Up

Luka: How do you build a business out of this technology in general aviation?

Alex List: The challenge in GA is that you look at it and think it's looking pretty good — aviation is 4% of the world economy. Then you look at 200,000 private pilots, 500,000 non-student pilots — the addressable market is like tens of millions of dollars. A venture capitalist needs to hear 10 to the 8th. But there are 450,000 remote pilots growing by 60,000 a year with a positive second derivative. Then you layer in counter-UAS and markets that are very underserved. There could be a real business here related to training and operations.

Jim: What are they willing to pay for? What's the value of having flown to every airport a thousand times?

Alex List: It's like what's the value of a Claude or ChatGPT subscription — you have access to the world's information, a thinking partner. We've taken that value and tailored it to a specific application: aviation. We've put a lot of work into making sure information is delivered in a way that's reliable and suitable to the context. There's the training aspect with partnerships like Sporty's, doing visual training and developing instrument scenarios. There's also the UAS space — delivering modern tactics to service members in barracks whose skills atrophy because of how quickly the Russia-Ukraine conflict is changing things. The way you build a business here is to have a moat related to data and have a moat related to the actual operation or provision of a particular service. Training is an excellent one, especially if you can allow other people to deliver training through what you do. That's becoming more and more of an emphasis for us.

Luka: Things that are less attractive from a business model perspective: starting with a tiny enthusiast user base; immediately going after some safety-critical dependence; requiring bespoke heavy integration with OEMs; broadly having OEM dependence without leverage. Those are pretty bad starting points. And also avoid the failure mode of becoming a services company in disguise — you start with genuinely interesting technology, but because of market fragmentation and slow regulation and the conservative buyer base, you end up being pulled into different bespoke work. 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.

Alex List: I was impressed by the Drone-Hand episode — their wedge is being an app on the DJI controller, able to bring machine learning to that app and scale to all these different farmers. That seems like a really excellent approach. I'm wondering what other approaches you guys have been impressed with, since you have pattern matching on aviation startup businesses than almost anybody else.

Luka: You want to see a clear operational pain that people are addressing and a recurring software-like revenue opportunity not requiring certification or SDCs. Come in with a narrow wedge that then has the potential to expand over time. You have some kind of data advantage that compounds with use.

Jim: What message would you want to send out to the entrepreneurial audience?

Alex List: You have to fundamentally build something people want. You need to take time and talk to your customers, talk to potential customers. A conversation with a customer is never a waste of time. The rest of my week is interacting with customers and building. One way or another, just go talk to people. You will go try to figure it out. It's a journey.