Are pilots going to be replaced by AI?

[Hank]

5 minutes ago, Will.iam said:

. . . one day a simi tractor trailer pulled out into the intersection and the tesla car saw it was clear to keep going but unfortunately the car’s cabin was too high to go under the trailer and thus ripped the top off the car and decapitated the driver. Computer systems are great until there is a situation that a programmer has not thought of and the code doesn’t have a contingency for. I’m sure they have added code or changed the camera angle to include hight requirements for the roof of the car to get through. 

So what else did they not think of???

[ilovecornfields]

Good thing no driver has ever been distracted or fallen asleep and hit the back of a tractor-trailer. I still remember the first fatal car accident I went to where a woman who appeared to have been in her early 20 reached over on the freeway to grab something from the back seat, smashed into the side of the guardrail, broke her neck and died on the spot. Eerie because when we got there she was still turned around with her hand under the seat but definitely dead. She’s probably be alive right now in a self-driving Tesla.

I completely agree about Gen Z’s view of tasks like driving. If my son could watch YouTube while going from PoInt A to Point B without being distracted by driving a car I’m sure he’d go for it. To be honest, I would too most of the time (not YouTube of course, but I’d love to get work done while I’m driving and have less to do at home). 
To the “computers aren’t perfect” crowd I’ll counter with “humans are less perfecter.”

[dzeleski]

2 hours ago, rbp said:

please explain in as much technical detail as possible what is non-deterministic about machine learning systems? in particular, where does the non-determinism originate? 

ML is not "AI" and neither are NLPs. They are all different topics within computer science and generally considered a subset under the topic of AI.

Here is a description of ML vs AI: https://ai.engineering.columbia.edu/ai-vs-machine-learning/ or https://azure.microsoft.com/en-us/resources/cloud-computing-dictionary/artificial-intelligence-vs-machine-learning/#introduction

Heres a description of what those two things mean: https://www.geeksforgeeks.org/difference-between-deterministic-and-non-deterministic-algorithms/

Hope that answers your question.

[dzeleski]

2 hours ago, N201MKTurbo said:

Can an AI system be validated? It seams it can change its mind. One aspect of software validation is you get the expected output every time. How could you prove that an AI system is reliable, when it can change its mind?

It seems that there is a lot of confusion between a traditionally software system, machine learning and AI. I'm not an expert on AI or machine learning, even though I just installed a machine learning system on my work computer. I try to avoid them because the licensing fees are quite high for what they do.

Yes but also not really. AI systems are validated using datasets. The issue with that is because data can vary so much its (at least at the time of writing this) basically impossible to test all the possible options an AI could execute upon. The issue with AI is that if you feed it the exact same data twice you probably wont get the same exact answer(there are memory training caveats here).

Sources on this are kinda hard to find because this is still kinda new but this article is the closest thing I could find for you: https://www.qed42.com/insights/perspectives/biztech/complete-guide-testing-ai-and-ml-applications

[N201MKTurbo]

17 minutes ago, dzeleski said:

Yes but also not really. AI systems are validated using datasets. The issue with that is because data can vary so much its (at least at the time of writing this) basically impossible to test all the possible options an AI could execute upon. The issue with AI is that if you feed it the exact same data twice you probably wont get the same exact answer(there are memory training caveats here).

Sources on this are kinda hard to find because this is still kinda new but this article is the closest thing I could find for you: https://www.qed42.com/insights/perspectives/biztech/complete-guide-testing-ai-and-ml-applications

One of the comments to the article points out that the tools referenced use AI to test software and they are not for testing AI software.

[dzeleski]

4 minutes ago, N201MKTurbo said:

One of the comments to the article points out that the tools referenced use AI to test software and they are not for testing AI software.

Yeah ignore the product promotion part. Tech sector is filled with "heres a problem, heres how you can pay us money to solve that problem"

[N201MKTurbo]

I have had people trying to sell me machine learning image analysis software for almost 20 years (Cognex, HALCON and others I can't remember) I tell them I will consider it if I can't solve the problem with standard image analysis methods. So far they have failed on every occasion. They would all like to sell you $10,000 license instead of the $2000 license. I'm very aware of where machine learning would be valuable, but I've never had a problem that required it.

[hais]

9 hours ago, rbp said:

please explain in as much technical detail as possible what is non-deterministic about machine learning systems? in particular, where does the non-determinism originate? 

When you write an algorithm, you prescribed exactly what should happen. When you train a ML model, you don't prescribe what should happen. Instead, you state whether a specific answer is acceptable.

Because you do not specify all acceptable answers, there's a chance that one that is not acceptable can be produced. Further, you do not know which answer will be generated.

This analogy below might help: 

Case 1: you program way points in your navigator. Then you turn on the auto pilot. You can always tell when which way the A/P will turn. This is similar to writing an algorithm.

Case 2: you tell your hanger neighbor that the hanger needs to be cleaned every Tuesday. On Tuesday, your neighbor flies out to a hamburger run. You give feedback that the flight was an acceptable solution. The Tuesday after, the neighbor parks their airplane in front of your hanger, blocking your flight. You did not anticipate this solution. Not only you don't understand why they picked this solution this time, but also you don't know which solution they will pick next.

Hopefully that explains.

 

 

[N201MKTurbo]

All machine learning systems I’ve ever worked with require that you give it a set of training images. Usually things that look like what you are looking for. Then optionally you give it a set of negative training images, that is things that don’t look like this. When it is all said and done, the machine learning software will pick objects out of a field of view and give them a score. You can specify that you only want objects that exceed some score. These things are very difficult to predict, you just have to play with it until you find a set of training images and scores that perform the way you want.

We built a machine once that sorted chili peppers. The customer wanted them all pointing the same way in the packaging. Machine Learning worked very well for this.

[Will.iam]

9 hours ago, ilovecornfields said:

Good thing no driver has ever been distracted or fallen asleep and hit the back of a tractor-trailer. I still remember the first fatal car accident I went to where a woman who appeared to have been in her early 20 reached over on the freeway to grab something from the back seat, smashed into the side of the guardrail, broke her neck and died on the spot. Eerie because when we got there she was still turned around with her hand under the seat but definitely dead. She’s probably be alive right now in a self-driving Tesla.

I completely agree about Gen Z’s view of tasks like driving. If my son could watch YouTube while going from PoInt A to Point B without being distracted by driving a car I’m sure he’d go for it. To be honest, I would too most of the time (not YouTube of course, but I’d love to get work done while I’m driving and have less to do at home). 
To the “computers aren’t perfect” crowd I’ll counter with “humans are less perfecter.”

To which i counter those computers are being programed by those less perfect people. Just like the the fighter jet pilot is the limiting factor, so too is the computer limited by the code of a person. If or when a computer can write better code than a human only then will it be better than a human for tasks that have high amounts of variables to adapt to. One of the biggest strengths humans have is the ability to adapt. Whether it’s a person that lost their hearing but adapt to reading lips or a person that has a stroke and recovers. Or the L-1011 that lost all three hydraulic systems that was thought to be impossible and yet the crew adapted to controlling it with what they had left to a controlled crash that allowed way more people to survive than thought possible. So yet computers are good for monitoring systems, playing chess or any known control group. But when the shit hits the fan, it’s a human that shines better than the computer. In flying it doesn’t happen as often as other modes of transportation but when things do go wrong in flying they tend to be more fatal than other modes of transportation. 

[T. Peterson]

On 3/25/2023 at 7:09 PM, 1980Mooney said:

Oh you mean like how the 3 pilots on Air France 447 handled the airspeed and angle of attack discrepancies on the 3 sensors? They had over 20,000 hours of experience between them. 

No thanks - I think the today's level of automation (AI) could do a better job than they did.   Certainly, the current level of GTP-3 Chatbot would work out the possible causes faster than they did (note: the pilots never did realize what was going on based on their statement from the voice recorder and their actions on the flight data recorder).   And I expect that there will be backup AI/Automation to the main automation - redundancies like in most aircraft systems.

And everyone is making comments about AI/automation capabilities by looking in the rearview mirror.  Let's face it - humans are at their peak capability (and many are below "peak").  AI/Automation is continually improving, becoming more sophisticated. Imagine what it will be in 5 years.

And BTW - soon your tax accountant with be AI. - it is just rules.  Probably a lot of lawyers will be replaced by AI - all the work junior lawyers do researching similar cases for strategy, precedent, etc.  AI will find more cases, more completely faster than any human.  And your accountant - and accountants at companies - will be AI.  It is just rules, ticking and tying.  And AI accountants wont embezzle from you.  White color jobs are at the greatest risk.  But AI can't and probably won't ever change your Depends when you are indigent, and bed ridden at a nursing home - those jobs are safe.

I think the argument can be made that the Air France pilots failed precisely because they were over dependent on Automation. They were great monitors, but not doers.

Precisely because of Air France and some other instances in which Pilots watched the automation fly the airplane into the ground there was a total paradigm shift towards more hands on flying. 
When our airline moved to the Airbus we stopped training stalls because the Bus “couldn’t stall.” After a couple of incidents at other airlines that was seen to be premature and we started training stalls again. 
Following a few more incidents like Air France, we now train not just stalls but full upset jet training.

It may be a little early to take your pilot uniforms to the thrift store.

[Pinecone]

6 hours ago, 1980Mooney said:

 

Why do you keep saying "once the plane was fully stalled the nose was about on the horizon".  The Final Report notes that the pitch was always 5 to 16 degrees positive (16 degrees up was the final data point at the moment of crash and 40% AOA).  The nose was never on the horizon (AH/AI) - it was always well above the horizon on the AH/AI.  The data in the FInal came from the Flight Recorder so it is the same data that the pilots saw.

And what is the point of your comments about "unless you have stalled a swept wing airplane would you be thinking stall" and " Swept wing aircraft to not behave in a stall like straight winged aircraft." ?  Are you suggesting that the pilots didn't know that?  - and that is why the PF didn't know how to recover from a stall?  Are you suggesting that the frail human pilot mind under stress reverted back to memories of learning to fly a straight wing C-172?  These are experienced professional pilots the are supposed to be able to be "thinking stall".  If humans can not be relied upon in situations like this then all the more reason to have more Automation and AI running the flight deck.

So, it was 5 degrees above the horizon.  That is much closer to the horizon that your Mooney would be with the yoke full back.

The point is, they did not recognize the stall because they never did one in the airplane and may not have done one in the sim.  Under Normal Law, an Airbus WON'T STALL.

[aviatoreb]

5 hours ago, Will.iam said:

To which i counter those computers are being programed by those less perfect people. Just like the the fighter jet pilot is the limiting factor, so too is the computer limited by the code of a person. If or when a computer can write better code than a human only then will it be better than a human for tasks that have high amounts of variables to adapt to. One of the biggest strengths humans have is the ability to adapt. Whether it’s a person that lost their hearing but adapt to reading lips or a person that has a stroke and recovers. Or the L-1011 that lost all three hydraulic systems that was thought to be impossible and yet the crew adapted to controlling it with what they had left to a controlled crash that allowed way more people to survive than thought possible. So yet computers are good for monitoring systems, playing chess or any known control group. But when the shit hits the fan, it’s a human that shines better than the computer. In flying it doesn’t happen as often as other modes of transportation but when things do go wrong in flying they tend to be more fatal than other modes of transportation. 

Computer code writers have an advantage that fighter pilots do not.  A single fighter pilot is asked to make a decision on a short time scale.  Sometimes a critical decision must occur in under a second.

Computer code writers have the complex task of writing robust codes to handle all such scenarios and clearly it is hard and errors can creep in all of sorts, from misunderstandings, straight up mistakes, oversights, syntax errors, bugs, and on and on.  But the computer programmer does get lots of time (relatively) to check, re-check and debug, AND teams of computer programmers have ways of cross checking and testing each other to diminish the many different kinds of errors.

[ilovecornfields]

10 hours ago, Will.iam said:

But when the shit hits the fan, it’s a human that shines better than the computer.

Why do you believe this to be true? Any evidence to support that statement? Even if it was true, the once in a lifetime event is somehow more significant than the daily human errors that occur and would have been prevented by a computer?

I’ve seen hundreds (if not thousands) of car accident victims. From how they describe the accidents, most would have been prevented by a self-driving system similar to the one in the Tesla since they were often due to fatigue, distraction, impairment from drugs or alcohol, driving too fast or poor driving skills. Perhaps some new ones would have occurred when the computer failed to see the semi-truck but these accidents are pretty rare compared to the rate of human-caused accidents. It still seems a lot of people don’t get this - to be better computers don’t have to be (and never will be) perfect. They just have to be better than humans (who aren’t that good to begin with). How many GA “pilot error” accidents would have been prevented by a computer? Running out of fuel, base to final stall/spin, VFR into IMC, CFIT, circle to land into mountains? Doubt a computer would have done that. How many gear-up landings would we see every week if this was done by a computer?

[Raymond J1]

80% of accidents are caused by human failure.
To replace the human with something else is to modify the occurrences on these 80% of human origin... But the machine is also fallible, so this does not balance the totality of human failures.
In an automatic machine, it is therefore necessary to resolve to accept a failure rate of about 20 - 25%, which will not necessarily lead to an accident.

[aviatoreb]

1 hour ago, ilovecornfields said:

Why do you believe this to be true? Any evidence to support that statement? Even if it was true, the once in a lifetime event is somehow more significant than the daily human errors that occur and would have been prevented by a computer?

I’ve seen hundreds (if not thousands) of car accident victims. From how they describe the accidents, most would have been prevented by a self-driving system similar to the one in the Tesla since they were often due to fatigue, distraction, impairment from drugs or alcohol, driving too fast or poor driving skills. Perhaps some new ones would have occurred when the computer failed to see the semi-truck but these accidents are pretty rare compared to the rate of human-caused accidents. It still seems a lot of people don’t get this - to be better computers don’t have to be (and never will be) perfect. They just have to be better than humans (who aren’t that good to begin with). How many GA “pilot error” accidents would have been prevented by a computer? Running out of fuel, base to final stall/spin, VFR into IMC, CFIT, circle to land into mountains? Doubt a computer would have done that. How many gear-up landings would we see every week if this was done by a computer?

Self driving certainly has its bugs and it is not ready for prime time in my opinion.  But it will be.  Humans make errors. The AI self driving makes errors.  The question is which makes errors at a higher rate.  Not can computers make zero errors.  There is no such thing as errorless, but there is such a thing as statistically safer.

Case in point.  Sitting on the coach causes all sorts of disease of sloth.  Running as an activity mitigates to a large degree those diseases but brings on its own problems for mortality.  Some people get run over as pedestrians while jogging, and that rarely happens when on the coach watching the price is right.  Some people get sudden heart attacks while running - which hard to say if that would have happened to the same person congenitally sooner or later with the life style of couch potato.  But in any case its a matter of choose your evil.

 

[T. Peterson]

6 hours ago, aviatoreb said:

Computer code writers have an advantage that fighter pilots do not.  A single fighter pilot is asked to make a decision on a short time scale.  Sometimes a critical decision must occur in under a second.

Computer code writers have the complex task of writing robust codes to handle all such scenarios and clearly it is hard and errors can creep in all of sorts, from misunderstandings, straight up mistakes, oversights, syntax errors, bugs, and on and on.  But the computer programmer does get lots of time (relatively) to check, re-check and debug, AND teams of computer programmers have ways of cross checking and testing each other to diminish the many different kinds of errors.

Grant it, but I think William’s point is that imperfect people will never design a perfect system. I agree.

[Hank]

11 minutes ago, aviatoreb said:

Self driving certainly has its bugs and it is not ready for prime time in my opinion.  But it will be.  Humans make errors. The AI self driving makes errors.  The question is which makes errors at a higher rate.  Not can computers make zero errors.  There is no such thing as errorless, but there is such a thing as statistically safer.

Case in point.  Sitting on the coach causes all sorts of disease of sloth.  Running as an activity mitigates to a large degree those diseases but brings on its own problems for mortality.  Some people get run over as pedestrians while jogging, and that rarely happens when on the coach watching the price is right.  Some people get sudden heart attacks while running - which hard to say if that would have happened to the same person congenitally sooner or later with the life style of couch potato.  But in any case its a matter of choose your evil.

Couch potatoes also rarely have shin splints, fewer knee / ankle / leg injuries. Runners typically have fewer weight-related issues, if they run far enough often enough. 

TANSTAAFL.

[There Ain't No Such Thing As A Free Lunch]

Everything has risks, try to pick the set of resulting risks that matches up with your expected input [effort / work / money]. Getting out of bed in the morning carries risk, but so does staying in bed all day . . . . .

[T. Peterson]

2 hours ago, ilovecornfields said:

Why do you believe this to be true? Any evidence to support that statement? Even if it was true, the once in a lifetime event is somehow more significant than the daily human errors that occur and would have been prevented by a computer?

I’ve seen hundreds (if not thousands) of car accident victims. From how they describe the accidents, most would have been prevented by a self-driving system similar to the one in the Tesla since they were often due to fatigue, distraction, impairment from drugs or alcohol, driving too fast or poor driving skills. Perhaps some new ones would have occurred when the computer failed to see the semi-truck but these accidents are pretty rare compared to the rate of human-caused accidents. It still seems a lot of people don’t get this - to be better computers don’t have to be (and never will be) perfect. They just have to be better than humans (who aren’t that good to begin with). How many GA “pilot error” accidents would have been prevented by a computer? Running out of fuel, base to final stall/spin, VFR into IMC, CFIT, circle to land into mountains? Doubt a computer would have done that. How many gear-up landings would we see every week if this was done by a computer?

I think Wall Street may have had a few computer glitches with pretty significant consequences.

Of course we know by experience the foibles of human pilots, but we are only imagining the perfect bliss of computer pilots. What we want to believe may not work out that perfectly.

[TheAv8r]

A driverless AI-driven car network will be the safest mode of transportation ever invented. Each car can talk to each other seamlessly, so it's more than just AI models to detect and account for driving variables, the network can literally know what each car is doing. It'd be like RVSM autopilots on steroids. 

If every car on an interstate or a highway was driven by AI as part of a driverless network, it would likely be not only the safest highway, but also the most efficient... no more crazy delays caused by that one car going 20 under the speed limit backing up a lane and causing everyone to try to get around it. Perfect distance spacing between each car to achieve an average speed that's the most efficient for the number of cars and lanes.

But therein lies the problem: every car. Outliers, the crazy drivers cutting across 4 lanes of traffic to make an exit by 30ft. The one guy manually driving going 50 above the speed limit weaving in and out of traffic, serving into lanes with 8 inches of space, etc. Those that refuse to be driven will introduce the instability in the network that AI has to accommodate for and right now, it's just not good enough to do that yet and public trust isn't there yet. Plus it would take decades to transition all the cars on the road to being AI autonomous cars, think about how many out there are still driving clunkers from the 90s and early 2000s. 

Where that goes... I don't know. Autonomous highways that only AI cars can drive is an expensive headache of an option. Making AI smart enough to accommodate for crazy driving variables (defensive driving) is another, but there will always be a degree of risk with that approach. 

[T. Peterson]

5 minutes ago, TheAv8r said:

A driverless AI-driven car network will be the safest mode of transportation ever invented. Each car can talk to each other seamlessly, so it's more than just AI models to detect and account for driving variables, the network can literally know what each car is doing. It'd be like RVSM autopilots on steroids. 

If every car on an interstate or a highway was driven by AI as part of a driverless network, it would likely be not only the safest highway, but also the most efficient... no more crazy delays caused by that one car going 20 under the speed limit backing up a lane and causing everyone to try to get around it. Perfect distance spacing between each car to achieve an average speed that's the most efficient for the number of cars and lanes.

But therein lies the problem: every car. Outliers, the crazy drivers cutting across 4 lanes of traffic to make an exit by 30ft. The one guy manually driving going 50 above the speed limit weaving in and out of traffic, serving into lanes with 8 inches of space, etc. Those that refuse to be driven will introduce the instability in the network that AI has to accommodate for and right now, it's just not good enough to do that yet and public trust isn't there yet. Plus it would take decades to transition all the cars on the road to being AI autonomous cars, think about how many out there are still driving clunkers from the 90s and early 2000s. 

Where that goes... I don't know. Autonomous highways that only AI cars can drive is an expensive headache of an option. Making AI smart enough to accommodate for crazy driving variables (defensive driving) is another, but there will always be a degree of risk with that approach. 

There are also factors beyond the mere technical development. What about people who enjoy driving for driving sake? The retired couple enjoying their convertible or the motorcyclist enjoying the curves. Or the young family that can’t afford an AI car? Not to worry, just bend to the collective will….hmmm……never been here before….

[aviatoreb]

21 minutes ago, TheAv8r said:

A driverless AI-driven car network will be the safest mode of transportation ever invented. Each car can talk to each other seamlessly, so it's more than just AI models to detect and account for driving variables, the network can literally know what each car is doing. It'd be like RVSM autopilots on steroids. 

If every car on an interstate or a highway was driven by AI as part of a driverless network, it would likely be not only the safest highway, but also the most efficient... no more crazy delays caused by that one car going 20 under the speed limit backing up a lane and causing everyone to try to get around it. Perfect distance spacing between each car to achieve an average speed that's the most efficient for the number of cars and lanes.

But therein lies the problem: every car. Outliers, the crazy drivers cutting across 4 lanes of traffic to make an exit by 30ft. The one guy manually driving going 50 above the speed limit weaving in and out of traffic, serving into lanes with 8 inches of space, etc. Those that refuse to be driven will introduce the instability in the network that AI has to accommodate for and right now, it's just not good enough to do that yet and public trust isn't there yet. Plus it would take decades to transition all the cars on the road to being AI autonomous cars, think about how many out there are still driving clunkers from the 90s and early 2000s. 

Where that goes... I don't know. Autonomous highways that only AI cars can drive is an expensive headache of an option. Making AI smart enough to accommodate for crazy driving variables (defensive driving) is another, but there will always be a degree of risk with that approach. 

I agree!  This is currently the most difficult environment which is AI driven cars must negotiate with irrational human driven cars.  I think this will be a brief period - what 30?  50 years?  When will it be that humans driving cars (and airplanes) are outlawed outright to be replaced 100% by computers.  The the computers can negotiate directly with each other.

[aviatoreb]

43 minutes ago, Hank said:

Couch potatoes also rarely have shin splints, fewer knee / ankle / leg injuries. Runners typically have fewer weight-related issues, if they run far enough often enough. 

TANSTAAFL.

[There Ain't No Such Thing As A Free Lunch]

Everything has risks, try to pick the set of resulting risks that matches up with your expected input [effort / work / money]. Getting out of bed in the morning carries risk, but so does staying in bed all day . . . . .

I agree!  But I didn't want to get into the ache's and pains and other injuries of sports - but instead I decide just to talk about mortality.

Couch potatoes get all sorts of aches and pains too.

Anyway - choose your evil or as you said, pick the set of resulting risks that ...

[ilovecornfields]

47 minutes ago, T. Peterson said:

Of course we know by experience the foibles of human pilots, but we are only imagining the perfect bliss of computer pilots. What we want to believe may not work out that perfectly.

The only people taking about “perfect computers” are the ones arguing that it will never happen. Of course computers aren’t perfect and their programing is imperfect. But it DOESN’T HAVE TO BE PERFECT TO BE BETTER THAN WHAT WE HAVE NOW. I apologize for the “screaming” but I think I’ve said the same thing almost 10 times now and keep seeing “but computers aren’t perfect…”

I think I’ll assign any further responses to Google Bard. Apparently it makes more sense than I do. And gets less frustrated.

[Will.iam]

30 minutes ago, ilovecornfields said:

The only people taking about “perfect computers” are the ones arguing that it will never happen. Of course computers aren’t perfect and their programing is imperfect. But it DOESN’T HAVE TO BE PERFECT TO BE BETTER THAN WHAT WE HAVE NOW. I apologize for the “screaming” but I think I’ve said the same thing almost 10 times now and keep seeing “but computers aren’t perfect…”

I think I’ll assign any further responses to Google Bard. Apparently it makes more sense than I do. And gets less frustrated.

I’m not arguing it will not happen, the question is when will it be safe enough to be accepted by the majority. And today is not the day.