4 ms·

Seems like a job a computer can do better. How soon before a corporation with "ai" takes over this role?

by markdown 4y ago

hgsgm 4y ago
The AI should be assisting the human ATCs by putting useful info on screen in order.
- j16sdiz 4y ago
  They tried —- it didn’t work well. It is hard to determine what is “useful” or “important”. In case of emergency or chaos, moving info around brings confusions.
refulgentis 4y ago
[flagged]
GalenErso 4y ago
You deserve more upvotes than you're getting, because your question is a fair one, and knowledge that may be obvious to some is going to be totally new to others. Some people would do well to remember that new people are constantly being born, and each person must learn everything from the start. I'm going to copy and paste a few answers from this 6 year old Reddit thread: https://old.reddit.com/r/explainlikeimfive/comments/5xqkxc/eli5_why_has_air_traffic_control_not_been/ https://old.reddit.com/r/explainlikeimfive/comments/5xqkxc/e... > Air traffic controllers think on the fly more than you would assume. They look at their radar scopes and plan each target to work with others by working their speed, altitude, and path. Computers could probably work out these, but then how would they relay that information to pilots? > Another thing to think about is the vast number of VFR aircraft who are not on a flight plan going A to B. ATC is able to communicate with them, learn their intentions "we saw something neat and would like to orbit here for 5 minutes", and help them see and avoid other aircraft. > The fact that they are humans who you can talk to is tremendously helpful in emergencies as well. If I'm fighting a sick airplane I might not have the time or focus to find the nearest suitable airport, get weather information and important notices, and plan a descent. A controller watching over me sure can though. Humans are flexible in a way computers can't be yet. ----- > All aviation-related technology is tightly regulated and goes through an immense amount of testing and consideration before use. > All of the obstacles we have around self-driving cars-- "whose fault is it if it crashes", "how do we know it's safe enough before letting it on the road", "do you still need a licensed driver", etc-- are 10x more problematic in the sky. It isn't necessarily harder to automate, but is harder to declare "officially safe enough". It's not enough for an automated ATC to be somewhat good or decent. There is little margin for error when the consequences of such an error could be the death of hundreds of people and the loss of a multimillion dollar plane. ATC in real life is much more difficult than it appears to casual observers.
- cardboardbach 4y ago
  [flagged]
  - refulgentis 4y ago
    This guys being too nice which leaves the door open to low effort replies: OP wasn’t even downvoted, and you have to watch about 10 min of air traffic videos on YouTube to understand. And I’m high on the GPT supply
  - bobthepanda 4y ago
    It may be smarter than the average human, but probably not smarter than trained humans specialized in ATC, at this particular time. ATC in particular depends on exact phrasing, in issuing commands, acknowledgement, and repetition back, so GPT’s ability to fudge something nice sounding is actually pretty bad.
    - flangola7 4y ago
      [flagged]
      - bobthepanda 4y ago
        ATC operation is not a search problem or an output generation problem, which is what GPT is primarily solving. It's the knapsack problem with a very complicated knapsack.
        flangola7 4y ago
        I don't know what knapsacks have to do with LLMs. I didn't say anything about search or output generation, nor did I say those are the only things transformer models can solve.
  - walterbell 4y ago
    > smarter than most humans In some pre-defined contexts.
jillesvangurp 4y ago
It's a matter of decades of tradition. They'd have to change the way this stuff works for this to work properly. If you think about it, 99% of ATC exchanges are entirely routine exchanges of very basic information. Requesting and answering such basic questions as: Who the hell are you, what type of airplane are you flying, what are your intentions, change course to X, altitude to Y, altimeter settings, etc. And that's before we start talking about more complicated approach and departure proceedings. Essentially all of that can be changed to exchanging information digitally and automatically. In fact most of that information exists digitally. It's not a technical challenge to do that. There are many ways to make this happen. The challenge is entirely institutional because it would represent massive changes to an industry that resists all form of change for decades on end. And that's for good reason: safety. But it also means that a lot of very flaky technology keeps on getting used and that pilots are task saturated with a ridiculous amount of work during safety critical portions of flight. When they get behind on their tasks, dangerous situations happen. When they stop practicing, they lose the ability to stay on top of things. It's that hard. And their pilot license shortly after (currency is not optional). If you'd design the procedures and communication from scratch using modern tools, you'd end up with a very different system than what we have today. The fun thing indeed is that with modern AI, what might happen instead is that we end up with AI assisted ATCs and pilots delegating the ritualistic exchange of information via VHF radios to AIs. They could just send each other digital messages instead that are then processed automatically and acknowledged manually when needed. But on the off chance of one of the pilots or atcs still being stuck in the nineteen sixties, it's going to be easier to pretend that everyone is and have AIs imitate human actors participating in this. So, text to speech and speech to text on both ends combined with an AI that has enough contextual information to take simple decisions and control systems escalating to pilots and controllers only when needed to might be more doable than coming up with a new system. I'd say chat gpt 4 would be well up for a "pretend you are an ATC controller" type role playing session.
- Out_of_Characte 4y ago
  I'm not sure how feasable AI or automation is. How would you respond to a failure in the automated system once there are more planes in the air than ATC controllers in that area? That could be a single point of failure. ATC Can optimise for severe degradation of service but how would automation be able to respond to a sometimes unavoidable impossible situation? You can define rules how planes should and shouldn't fly, verifying that this is always possible would be a nightmare.
jackweirdy 4y ago
We have a way already to keep routine information off the radios - CPDLC - https://en.m.wikipedia.org/wiki/Controller%E2%80%93pilot_data_link_communications https://en.m.wikipedia.org/wiki/Controller%E2%80%93pilot_dat... Entirely possible we could find more things to move there with humans

Seems like a job a computer can do better. How soon before a corporation with "ai" takes over this role?

hgsgm 4y ago

The AI should be assisting the human ATCs by putting useful info on screen in order.

j16sdiz 4y ago
They tried —- it didn’t work well. It is hard to determine what is “useful” or “important”. In case of emergency or chaos, moving info around brings confusions.

refulgentis 4y ago

[flagged]

GalenErso 4y ago

You deserve more upvotes than you're getting, because your question is a fair one, and knowledge that may be obvious to some is going to be totally new to others. Some people would do well to remember that new people are constantly being born, and each person must learn everything from the start. I'm going to copy and paste a few answers from this 6 year old Reddit thread: https://old.reddit.com/r/explainlikeimfive/comments/5xqkxc/eli5_why_has_air_traffic_control_not_been/ https://old.reddit.com/r/explainlikeimfive/comments/5xqkxc/e... > Air traffic controllers think on the fly more than you would assume. They look at their radar scopes and plan each target to work with others by working their speed, altitude, and path. Computers could probably work out these, but then how would they relay that information to pilots? > Another thing to think about is the vast number of VFR aircraft who are not on a flight plan going A to B. ATC is able to communicate with them, learn their intentions "we saw something neat and would like to orbit here for 5 minutes", and help them see and avoid other aircraft. > The fact that they are humans who you can talk to is tremendously helpful in emergencies as well. If I'm fighting a sick airplane I might not have the time or focus to find the nearest suitable airport, get weather information and important notices, and plan a descent. A controller watching over me sure can though. Humans are flexible in a way computers can't be yet. ----- > All aviation-related technology is tightly regulated and goes through an immense amount of testing and consideration before use. > All of the obstacles we have around self-driving cars-- "whose fault is it if it crashes", "how do we know it's safe enough before letting it on the road", "do you still need a licensed driver", etc-- are 10x more problematic in the sky. It isn't necessarily harder to automate, but is harder to declare "officially safe enough". It's not enough for an automated ATC to be somewhat good or decent. There is little margin for error when the consequences of such an error could be the death of hundreds of people and the loss of a multimillion dollar plane. ATC in real life is much more difficult than it appears to casual observers.

cardboardbach 4y ago
[flagged]
- refulgentis 4y ago
  This guys being too nice which leaves the door open to low effort replies: OP wasn’t even downvoted, and you have to watch about 10 min of air traffic videos on YouTube to understand. And I’m high on the GPT supply
- bobthepanda 4y ago
  It may be smarter than the average human, but probably not smarter than trained humans specialized in ATC, at this particular time. ATC in particular depends on exact phrasing, in issuing commands, acknowledgement, and repetition back, so GPT’s ability to fudge something nice sounding is actually pretty bad.
  - flangola7 4y ago
    [flagged]
    - bobthepanda 4y ago
      ATC operation is not a search problem or an output generation problem, which is what GPT is primarily solving. It's the knapsack problem with a very complicated knapsack.
      - flangola7 4y ago
        I don't know what knapsacks have to do with LLMs. I didn't say anything about search or output generation, nor did I say those are the only things transformer models can solve.
- walterbell 4y ago
  > smarter than most humans In some pre-defined contexts.

jillesvangurp 4y ago

It's a matter of decades of tradition. They'd have to change the way this stuff works for this to work properly. If you think about it, 99% of ATC exchanges are entirely routine exchanges of very basic information. Requesting and answering such basic questions as: Who the hell are you, what type of airplane are you flying, what are your intentions, change course to X, altitude to Y, altimeter settings, etc. And that's before we start talking about more complicated approach and departure proceedings. Essentially all of that can be changed to exchanging information digitally and automatically. In fact most of that information exists digitally. It's not a technical challenge to do that. There are many ways to make this happen. The challenge is entirely institutional because it would represent massive changes to an industry that resists all form of change for decades on end. And that's for good reason: safety. But it also means that a lot of very flaky technology keeps on getting used and that pilots are task saturated with a ridiculous amount of work during safety critical portions of flight. When they get behind on their tasks, dangerous situations happen. When they stop practicing, they lose the ability to stay on top of things. It's that hard. And their pilot license shortly after (currency is not optional). If you'd design the procedures and communication from scratch using modern tools, you'd end up with a very different system than what we have today. The fun thing indeed is that with modern AI, what might happen instead is that we end up with AI assisted ATCs and pilots delegating the ritualistic exchange of information via VHF radios to AIs. They could just send each other digital messages instead that are then processed automatically and acknowledged manually when needed. But on the off chance of one of the pilots or atcs still being stuck in the nineteen sixties, it's going to be easier to pretend that everyone is and have AIs imitate human actors participating in this. So, text to speech and speech to text on both ends combined with an AI that has enough contextual information to take simple decisions and control systems escalating to pilots and controllers only when needed to might be more doable than coming up with a new system. I'd say chat gpt 4 would be well up for a "pretend you are an ATC controller" type role playing session.

Out_of_Characte 4y ago
I'm not sure how feasable AI or automation is. How would you respond to a failure in the automated system once there are more planes in the air than ATC controllers in that area? That could be a single point of failure. ATC Can optimise for severe degradation of service but how would automation be able to respond to a sometimes unavoidable impossible situation? You can define rules how planes should and shouldn't fly, verifying that this is always possible would be a nightmare.

jackweirdy 4y ago

We have a way already to keep routine information off the radios - CPDLC - https://en.m.wikipedia.org/wiki/Controller%E2%80%93pilot_data_link_communications https://en.m.wikipedia.org/wiki/Controller%E2%80%93pilot_dat... Entirely possible we could find more things to move there with humans