Sikorsky's story is really fascinating to me. 4 years after immigrating from Russia he founded an aviation company in USA. And in 15 years he was already building airplanes and then first ever helicopter! (https://upload.wikimedia.org/wikipedia/commons/2/27/Igor_Sik...)
If he stayed in Russia the course of human history might have been very different ... if he was developing all that technology for the USSR... and not USA.
tell me about boot strapping!! :
"In 1923, Sikorsky formed the Sikorsky Manufacturing Company in Roosevelt, New York.[36] He was helped by several former Russian military officers. Among Sikorsky's chief supporters was composer Sergei Rachmaninoff, who introduced himself by writing a check for US$5,000 (approximately $61,000 in 2007).[37] Although his prototype was damaged in its first test flight, Sikorsky persuaded his reluctant backers to invest another $2,500. With the additional funds, he produced the S-29, one of the first twin-engine aircraft in America, with a capacity for 14 passengers and a speed of 115 mph.[38] The performance of the S-29, slow compared to military aircraft of 1918, proved to be a "make or break" moment for Sikorsky's funding."
wow I did not know he was Russian, thats quite a feat.
also fascinating is Mazda...the founder literally built it on ground zero in Hiroshima, using unexploded US munitions and whatever parts he could scavenge.
It's like some people....they see adversity...and they make it their bitch. its awe inspiring.
He probably would have died in the gulag or at the very least not achieved what he did. Stalin was not fond of the aerospace industry so the guy who's at the leading edge of designing rotary winged aircraft would likely find his name high up on the wrong list and if he escaped that fate the money wouldn't have exactly been flowing his way.
>Its definitely a strong possibility that his talent would have gone unrealized!
Mikhail Kalashnikov wanted to design agricultural equipment because he wanted to help prevent famine. Who knows if he would have been any good at it but that door closed on June 22 1941.
It's hard to say what would have happened had history been different.
> If he stayed in Russia the course of human history might have been very different... if he was developing all that technology for the USSR... and not USA.
I have serious doubts about this. His talents might be lost for humanity completely. Post-revolution and before WW2 there were a lot of constructors and engineers who were prosecuted, displaced or even worse by the Soviets.
The wikipedia says "After the Bolshevik revolution began in 1917, Igor Sikorsky fled his homeland, because the new government threatened to shoot him." (the citation link is broken, unfortunately).
This reminds me of Sergey Korolev's history. He was basically the father of Soviet rocket-building. He has survived the purges by pure luck. The leaders of his institute were executed, he was tortured to get the confession, lost his teeth, went through the gulag and got back. How many of these talented people ended up being not so lucky is hard to imagine.
Even the work results might not have protected them. For example, take the chief engineer of T-34 engine https://en.wikipedia.org/wiki/Konstantin_Chelpan. He got awarded for the invention, arrested and executed the next year, and then rehabilitated a few years later.
There were so many stories like these.
Similar also to the aftermath of the Cultural Revolution in China, many scientists and other intellectuals never came back from those camps[1]. Presumably modern China would look quite different if that hadn't been the case.
The downfall of the Russian Empire was really good for American aviation. Other examples are Seversky who founded Republic aviation and Kartveli who designed many of the important aircraft for Seversky were both Russian noblemen who would have been killed by the Bolsheviks.
Urban air mobility space is heating up. Blade and Sikorsky's AAG announced an agreement to provide a new on-demand urban mobility option to the NYC area. Uber has their Elevate team working on this problem as well. Blade launched an urban mobility pilot program in the Bay Area last week. Received an early, invite code from a friend last week, SF-3FTPTV.
>> Flying a helicopter requires such an intense mental load, he says, that even small things like pushing the “talk” button can put a novice pilot in peril.
No they are not that hard to fly. A bare-bones helicopter with nothing more than the minimum parts to qualify as a helicopter is indeed an unsteady beast. But such helicopters are rare, used mostly for training purposes. Modern machines, even ancient ones, have things like gyroscopes to take much of the load off the pilot. And autopilots really do work (auto, not autonomous). They can hold a steady heading/alt. Feed them data from a radar altimeter and they can hover like a rock.
The computers can fly the aircraft, they can literally make it move as needed, but that is a totally different problem than deciding where to move the aircraft. The pilot's job is making the judgement calls necessary to keep the aircraft safe. Show me a computer than can determine whether an approach is safe enough to execute, whether the weather en route is acceptable. Show me a computer than an judge which path to take to avoid carrying an unsteady slung load over someone's head.
Well, "rock" as in solid-as-as. Watch any of the new rescue helos hovering over a ship.
There are legends in AF helo communities about entire crews falling asleep during long hovers. (The anti-submarine helicopters have to hover in place while dipping their sonar into the water.)
This is really interesting and surprising to me. Can any helicopter pilot explain why hovering is not as easy as just taking your hands off the controls? That's how I always imagined a helicopter would work!
It's like standing on a basketball, any perturbation in the wrong direction causes it to accelerate in that direction. The whole thing is vibrating in multiple planes, so it requires constant control inputs to keep it from veering off.
Ideally you can trim a modern helicopter out much like you would an airplane however it's still a dynamically (?) unstable platform. I'm far removed from aero and nobody really thinks about the details of that stuff when they're flying. Here's a good summary:
A hover isn't as dynamic as forward flight. A fixed-wing aircraft like a glider can self-regulate using aerodynamic forces. For instance: If the nose tips up, the plane slows. The wings generate less lift and the nose tips back down. Plane then speeds up ... etc. But a hovering helicopter doesn't experience any significant airflow changes. If it's nose tips up it just keeps tipping up. The helo then starts sliding backwards until the tail catches the wind, spinning everything 180 like a weather vain and very quickly you aren't flying any more. The pilot/computer/gyro needs to keep ahead of these forces.
I'm pretty sure everything in your last paragraph could be done with enough good input (sensors, approach mins, DTED, Digital Wx updates, etc) and the correct application of 1's and 0's. Not really that hard of a problem to solve unless you're operating somewhere very data-austere.
To your first point-and I think you hit it but I wanted to emphasize-some helicopters are hard to fly, some mission sets are more difficult than others; hovering with a 30 knot tailwind in a $32 million dollar aircraft with SCAS and AFCS is still a pain in the ass especially when it's low light and you're on goggles...but I guess the AI wouldn't need to be aided, so call that a win.
>> ... everything in your last paragraph could be done with enough good input
Lol. Helicopters are not 747s. They operate very locally and do things that are not as generic as approaching SFO in marginal weather. A helicopter has to change its operations in reaction not to measured weather, but to the specifics of individual trees. Show me a computer that can take the input: "That tree looks like it is about to fall over" and decide whether to continue a rescue or give up and negotiate a new approach. SAR pilots do that daily.
Sorry just being short on a comment that starts with lol. Look at for instance SAR mapping. I can have a “certain curr gen aircraft” SAR map a zone or AO for me and tell me A LOT (UNCLAS/open source).
I appreciate the appeal to patriotism buddy but I’m right there with you in a flight suit. The fact that you are in a job that requires a flight suit says a lot less about your experience than your comments have. I don’t like to throw the word boot out but...
I suspect you have not tried hovering in a helicopter.
> autopilots really do work
There are actually very few autopilots that can hover. Not just because it's technically hard (the soviet Kamov helicopters could maintain hover back in 1980's), but also because of the certification requirements.
If a DJI drone hits a tree it's not a big deal. The story is very different when a helicopter does it.
> But SARA doesn’t rely as much on “high-order” functions like AI since those are harder to certify. (“To be FAA certifiable, you need a certain level of determinism in the outcomes,” Van Buiten says.)
i've just recently finished peter watts "freeze frame revolution" of the sunflower cycle, where the gate-building spaceship eriophora is mostly on its own for thousands of years while the crew sleeps, guided by an AI called "the chimp".
-- spoilers --
one of the key ideas in this book is that the ships AI - the chimp - is not advanced at all, with a synapse count of roughly the chimp. anything more intelligent would get unstable and develop its own motives, so the original builders constructed it to be comparatively dumb and thus stable, predictable and deterministic (the humans are woken up only in case something unexpected happens which requires more creativity and brainpower).
I think Sikorsky's probably ahead in the race for realistic autonomous flight. Sikorsky isn't a trendy/hot company by any means, but they're reliable, they deliver, and they have the contact infrastructure with FAA and NTSB and other government certification orgs that would handle this new development. I don't know if they'd offer "air-taxis", but decreasing the cost of helicopter travel so that a seed/Series A startup CEO can travel vs. the Fortune 500 CEOs would be a major improvement in and of itself.
Reminds me of a gag in Bojack Horseman. The scene is a metropole traffic jam, with the cars standing still in a neat column, and in the background you can see an analogous column of helicopters, hovering still in the sky. It's a helicopter jam.
I don't think the pilot factors in that much into the cost of operating a helicopter. Even a small helicopter (e.g. Robinson R-44 which is a bare bones trainer) needs a ~250k complete overhaul every 2200 hours/12 years in service. Add $60 dollars per hour in fuel and 60-100 per hour of miscellaneous other maintenance, and your variable costs quickly approach 300/hr. Add in all the fixed overhead, and a tiny helicopter that can take 2 passengers and limited baggage will run $400 dollars per hour easily (which is about what they rent for). A bigger turbine helicopter is significantly more expensive.
Most of this is because helicopters are complex machines (you're spinning really big blades pretty fast. Then using the blades + bearing to also lift the weight of the helicopter + more. Oh and you're changing the pitch angle of the blades. Oh you mean changing the pitch angles of the blades WHILE THEY GO AROUND...) and any failure in any of these parts is usually fatal so they have to be built and maintained to a very high degree of reliability.
Airplanes are much simpler (the spinning propeller attached to an engine is one part. The wing generating lift is another part. the flight controls are yet another part) and have more opportunities for redundancy ( a wing has multiple spars, and is attached to the airplane with many bolts. All helicopter blades meet in one hub, which is attached on one axis).
Combine that with aircraft scaling up more (you can build 500 person aircraft, but only 20 person or so helicopters), going much faster (a 120 dollar/hr propeller plane will outrun many/most helicopters) and the cost to go a given distance by plane will always be much cheaper than a helicopter.
Nah, helicopters are relatively simple machines, literally every person in this site can understand the basics of helicopter technology (of course, getting proficient and enabling a career in heli-tech is a whole different story). The only difficulty is making them bulletproof, but nowadays the crashes due to technical problems are rare.
Also, helicopters are generally used for other purposes than planes. Transporting thousands of men thousands of kilometers is probably not a job for helicopters but that does not imply that helicopters are useless.
I'm less comparing the helicopter approach with airplanes, and more comparing them to multirotors, where I think a lot of the autonomous flight hype is coming from, and where the use cases substitute rather than complement each other. In that sense, the fixed costs are less "baked-in" to the product. You can swap out helicopter blades for better composites, and add in a glass cockpit, but the fundamental control system still holds. You can't really change the number of motors in a multicopter without a full-scale overhaul of the avionics suite. And I sure hope the FAA wouldn't let a multicopter with some number of failed motors take off just because it doesn't immediately fall out of the sky.
I would agree with you operating a helicopter will always be expensive. I don't anticipate autonomous helicopters to be price competitive with something like cars. People pay the premium for vertical flight to gain the benefits of vertical flight. In addition, I'm pretty sure that given the high fixed price, that market would rather pay more to get a premium product rather than accept an inferior product (e.g. insufficient range, speed, flight time, safety, etc). I was more saying making helicopter flights marginally cheaper will make said flights marginally more available to more people.
There might be game changers down the line. For example, Sikorsky has a lot of experience with experimental control systems like hybrid helicopters that may reduce operational expenses, and if they can prove SARA/derivatives are as reliable or better than a human pilot and convince the general public/unions to fly without a pilot, they might design helicopters that are more maintenance-oriented. But as with all things, it's more important to make sure new innovations are deliverable and provably progressive.
> But SARA doesn’t rely as much on “high-order” functions like AI since those are harder to certify. (“To be FAA certifiable, you need a certain level of determinism in the outcomes,” Van Buiten says.)
I wonder whether a hybrid approach is the future, since in some tasks neural networks are just far better than anything we have. If the only tasks of the neural network is to estimate/classify some sensor-input and is trained in a purely supervised setting, the "right thing" for the neural network is still pretty well defined and rigorous testing should be possible (simple tasks can be very complex to implement). Then, interpretable, high-level reasoning could be solved by old-school coding (and maybe verifying).
This is not possible with end-to-end training.
But I am not sure what they mean, normally neural network (and their training) is purely deterministic. It's not that they are just very good at rolling a dice.
I am not into this stuff (autonomous, "intelligent" systems, more the data-analysis guy), but I would use neural networks for simple to define, hard problems that involve a lot of noisy data (where some kind of accuracy on some test-set is a well-defined metric) and then build a higher-level reasing system by hand.
I think they are using the wrong word here, I think by "determinism" they mean interpretable. You train your big neural network to classify images, it works fine on the test set but then one day in the wild you discover your network classified a turtle as a rifle [0]. Why did it do that? what can we change to fix that, how could we have seen this coming? Those answers won't come easily compared to a more old school system.
I am reasonably well versed in the whole adversarial-input research. I am not sure whether this is really the problem. It's a tradeoff. This essentially introduces a random chance that you classification will fail (since adversarial input is not constrained to hard-to-classify situations), but while less accurate methods may be more robust to random errors (which is hard to verify for many approaches) but they have a higher likelihood to fail with hard situations.
Doesn't every component in an aircraft have a random chance of failure? There's even a name for one critical component, the Jesus nut.
I don't really get the interpretable argument. I think what you want is to verify it to a reasonable degree.
> Doesn't every component in an aircraft have a random chance of failure? There's even a name for one critical component, the Jesus nut.
Two things;
1) This might be theoretically true, but Boeing and other aircraft manufacturers monitor data on every component at a very fine-grained level and they use inspections to predict what will go bad and when. They're even adding real-time capabilities to it so that parts are replaced immediately after a flight if imminent failure is predicted; https://www.boeing.com/commercial/aeromagazine/articles/qtr_...
2) The core issue over here isn't that a component can fail. The issue is that when it fails can we trace the error and forestall it in the future? If the error is something that's unreplicable and unpredictable then that's quite literally impossible and it makes the machine by its very definition untenably unsafe as you don't know when or why it won't work.
Someday neural networks will become a part of safety-critical systems, but this generation of neural networks probably won't be it.
I convinced that neural network will never be interpretable like code. They will just get even more complex, so to employ them in critical applications (my motivation is more healthcare, where similiar constraints are present. People can die if there's a wrong prediction) we have to build a notion why we allow something and what we don't allow without some human in the loop getting convined that the reasoning itself is sound. I think this boils down to assumptions about robustness (like the l2-norm adversarial examples in image classification) and statistics.
to 1). Neural networks are code, they are always 100% the same if you ship it. So you can test them way more thoughly than some equipment, which (i hope) would allow you to minimze the chance of failure to a reasonable degree.
to 2) "you don't know when or why it won't work." I think a fundamental disagreement lies here. I am convinced that this is not possible with these complex function approximators, where we just optimize them until we are happy. If we adopt a thinking like above, we can retrain the networks with these additional inputs, adjust our notion of robustness and think whether there's a flaw in our statistical assumptions.
But these are just random thoughts. I am in touch with people working in healthcare with this stuff, so there's some exposure to these kinds of problems. But I have never read any work discussing these issues or really reflection whether my reasoning is actually sound.
Maybe I'm wrong and we can somehow make them interpretable in a sense that is actually relevant to verifying them.
I'm impressed with what they've been able to do, and happy to see that they're developing in-house. When I was an intern there in the mid-2000's they were still working on fly-by-wire and most (all?) of their programming was done by a separate company.
41 comments
[ 2.7 ms ] story [ 77.3 ms ] threadhttps://news.ycombinator.com/item?id=19318425
If he stayed in Russia the course of human history might have been very different ... if he was developing all that technology for the USSR... and not USA.
tell me about boot strapping!! :
"In 1923, Sikorsky formed the Sikorsky Manufacturing Company in Roosevelt, New York.[36] He was helped by several former Russian military officers. Among Sikorsky's chief supporters was composer Sergei Rachmaninoff, who introduced himself by writing a check for US$5,000 (approximately $61,000 in 2007).[37] Although his prototype was damaged in its first test flight, Sikorsky persuaded his reluctant backers to invest another $2,500. With the additional funds, he produced the S-29, one of the first twin-engine aircraft in America, with a capacity for 14 passengers and a speed of 115 mph.[38] The performance of the S-29, slow compared to military aircraft of 1918, proved to be a "make or break" moment for Sikorsky's funding."
also fascinating is Mazda...the founder literally built it on ground zero in Hiroshima, using unexploded US munitions and whatever parts he could scavenge.
It's like some people....they see adversity...and they make it their bitch. its awe inspiring.
Imagine it didn't. Literally the outcome of WWII could have been very different.
Of course "what if's" dont count. Still fascinating to me how much he's accomplished 100 years ago!
Here we are today burning 100's of millions in cash of investors money with "0 results" (really). lol
Mikhail Kalashnikov wanted to design agricultural equipment because he wanted to help prevent famine. Who knows if he would have been any good at it but that door closed on June 22 1941.
It's hard to say what would have happened had history been different.
I have serious doubts about this. His talents might be lost for humanity completely. Post-revolution and before WW2 there were a lot of constructors and engineers who were prosecuted, displaced or even worse by the Soviets.
The wikipedia says "After the Bolshevik revolution began in 1917, Igor Sikorsky fled his homeland, because the new government threatened to shoot him." (the citation link is broken, unfortunately).
This reminds me of Sergey Korolev's history. He was basically the father of Soviet rocket-building. He has survived the purges by pure luck. The leaders of his institute were executed, he was tortured to get the confession, lost his teeth, went through the gulag and got back. How many of these talented people ended up being not so lucky is hard to imagine.
Even the work results might not have protected them. For example, take the chief engineer of T-34 engine https://en.wikipedia.org/wiki/Konstantin_Chelpan. He got awarded for the invention, arrested and executed the next year, and then rehabilitated a few years later. There were so many stories like these.
1: https://en.wikipedia.org/wiki/Cultural_Revolution#Education
https://en.wikipedia.org/wiki/Alexander_P._de_Seversky
https://en.wikipedia.org/wiki/Alexander_Kartveli
No they are not that hard to fly. A bare-bones helicopter with nothing more than the minimum parts to qualify as a helicopter is indeed an unsteady beast. But such helicopters are rare, used mostly for training purposes. Modern machines, even ancient ones, have things like gyroscopes to take much of the load off the pilot. And autopilots really do work (auto, not autonomous). They can hold a steady heading/alt. Feed them data from a radar altimeter and they can hover like a rock.
The computers can fly the aircraft, they can literally make it move as needed, but that is a totally different problem than deciding where to move the aircraft. The pilot's job is making the judgement calls necessary to keep the aircraft safe. Show me a computer than can determine whether an approach is safe enough to execute, whether the weather en route is acceptable. Show me a computer than an judge which path to take to avoid carrying an unsteady slung load over someone's head.
That's not good, right?
There are legends in AF helo communities about entire crews falling asleep during long hovers. (The anti-submarine helicopters have to hover in place while dipping their sonar into the water.)
> The ships hung in the sky in much the same way that bricks don't.
https://aviation.stackexchange.com/questions/35764/are-helic...
To your first point-and I think you hit it but I wanted to emphasize-some helicopters are hard to fly, some mission sets are more difficult than others; hovering with a 30 knot tailwind in a $32 million dollar aircraft with SCAS and AFCS is still a pain in the ass especially when it's low light and you're on goggles...but I guess the AI wouldn't need to be aided, so call that a win.
Lol. Helicopters are not 747s. They operate very locally and do things that are not as generic as approaching SFO in marginal weather. A helicopter has to change its operations in reaction not to measured weather, but to the specifics of individual trees. Show me a computer that can take the input: "That tree looks like it is about to fall over" and decide whether to continue a rescue or give up and negotiate a new approach. SAR pilots do that daily.
Considering the way you lambaste ideas that it is becoming increasingly evident you don’t fully understand, I’d be interested in your credentials.
I suspect you have not tried hovering in a helicopter.
> autopilots really do work
There are actually very few autopilots that can hover. Not just because it's technically hard (the soviet Kamov helicopters could maintain hover back in 1980's), but also because of the certification requirements.
If a DJI drone hits a tree it's not a big deal. The story is very different when a helicopter does it.
i've just recently finished peter watts "freeze frame revolution" of the sunflower cycle, where the gate-building spaceship eriophora is mostly on its own for thousands of years while the crew sleeps, guided by an AI called "the chimp".
-- spoilers --
one of the key ideas in this book is that the ships AI - the chimp - is not advanced at all, with a synapse count of roughly the chimp. anything more intelligent would get unstable and develop its own motives, so the original builders constructed it to be comparatively dumb and thus stable, predictable and deterministic (the humans are woken up only in case something unexpected happens which requires more creativity and brainpower).
Most of this is because helicopters are complex machines (you're spinning really big blades pretty fast. Then using the blades + bearing to also lift the weight of the helicopter + more. Oh and you're changing the pitch angle of the blades. Oh you mean changing the pitch angles of the blades WHILE THEY GO AROUND...) and any failure in any of these parts is usually fatal so they have to be built and maintained to a very high degree of reliability.
Airplanes are much simpler (the spinning propeller attached to an engine is one part. The wing generating lift is another part. the flight controls are yet another part) and have more opportunities for redundancy ( a wing has multiple spars, and is attached to the airplane with many bolts. All helicopter blades meet in one hub, which is attached on one axis).
Combine that with aircraft scaling up more (you can build 500 person aircraft, but only 20 person or so helicopters), going much faster (a 120 dollar/hr propeller plane will outrun many/most helicopters) and the cost to go a given distance by plane will always be much cheaper than a helicopter.
Also, helicopters are generally used for other purposes than planes. Transporting thousands of men thousands of kilometers is probably not a job for helicopters but that does not imply that helicopters are useless.
I would agree with you operating a helicopter will always be expensive. I don't anticipate autonomous helicopters to be price competitive with something like cars. People pay the premium for vertical flight to gain the benefits of vertical flight. In addition, I'm pretty sure that given the high fixed price, that market would rather pay more to get a premium product rather than accept an inferior product (e.g. insufficient range, speed, flight time, safety, etc). I was more saying making helicopter flights marginally cheaper will make said flights marginally more available to more people.
There might be game changers down the line. For example, Sikorsky has a lot of experience with experimental control systems like hybrid helicopters that may reduce operational expenses, and if they can prove SARA/derivatives are as reliable or better than a human pilot and convince the general public/unions to fly without a pilot, they might design helicopters that are more maintenance-oriented. But as with all things, it's more important to make sure new innovations are deliverable and provably progressive.
I wonder whether a hybrid approach is the future, since in some tasks neural networks are just far better than anything we have. If the only tasks of the neural network is to estimate/classify some sensor-input and is trained in a purely supervised setting, the "right thing" for the neural network is still pretty well defined and rigorous testing should be possible (simple tasks can be very complex to implement). Then, interpretable, high-level reasoning could be solved by old-school coding (and maybe verifying).
This is not possible with end-to-end training.
But I am not sure what they mean, normally neural network (and their training) is purely deterministic. It's not that they are just very good at rolling a dice.
I am not into this stuff (autonomous, "intelligent" systems, more the data-analysis guy), but I would use neural networks for simple to define, hard problems that involve a lot of noisy data (where some kind of accuracy on some test-set is a well-defined metric) and then build a higher-level reasing system by hand.
[0] https://arxiv.org/abs/1707.07397
Doesn't every component in an aircraft have a random chance of failure? There's even a name for one critical component, the Jesus nut.
I don't really get the interpretable argument. I think what you want is to verify it to a reasonable degree.
Two things;
1) This might be theoretically true, but Boeing and other aircraft manufacturers monitor data on every component at a very fine-grained level and they use inspections to predict what will go bad and when. They're even adding real-time capabilities to it so that parts are replaced immediately after a flight if imminent failure is predicted; https://www.boeing.com/commercial/aeromagazine/articles/qtr_...
2) The core issue over here isn't that a component can fail. The issue is that when it fails can we trace the error and forestall it in the future? If the error is something that's unreplicable and unpredictable then that's quite literally impossible and it makes the machine by its very definition untenably unsafe as you don't know when or why it won't work.
Someday neural networks will become a part of safety-critical systems, but this generation of neural networks probably won't be it.
to 1). Neural networks are code, they are always 100% the same if you ship it. So you can test them way more thoughly than some equipment, which (i hope) would allow you to minimze the chance of failure to a reasonable degree.
to 2) "you don't know when or why it won't work." I think a fundamental disagreement lies here. I am convinced that this is not possible with these complex function approximators, where we just optimize them until we are happy. If we adopt a thinking like above, we can retrain the networks with these additional inputs, adjust our notion of robustness and think whether there's a flaw in our statistical assumptions.
But these are just random thoughts. I am in touch with people working in healthcare with this stuff, so there's some exposure to these kinds of problems. But I have never read any work discussing these issues or really reflection whether my reasoning is actually sound.
Maybe I'm wrong and we can somehow make them interpretable in a sense that is actually relevant to verifying them.