- cross-posted to:
- [email protected]
- cross-posted to:
- [email protected]
You must log in or register to comment.
The fact that anyone is willing to set foot on an airplane in the United States with the current administration running things is astounding to me.
I mean, it’s still statistically safer than driving, and will be until like the fifth major plane crash.
My wife keeps talking about wanting to fly up to New England this year, but I’m not setting foot in an airport until I po longer run the risk of being sent to a concentration camp.
But, this program has nothing to do with LLMs or GenAI. It’s just good ol’ fashioned machine learning.
It’s much less safe than it was and it’s fucked up, it is still monumentally safer than driving. People are shit at risk assessment when it comes to air travel.
I don’t drive either though, I cycle to get around. Which is harder to tell how dangerous it is because the health benefits are usually greater than the risks.
The U.S. still takes aerospace safety very seriously, which is why it’s so safe.
As if i needed another reason not to fly
I avoid it as much as humanly possible. Haven’t flown in years. And it looks so much worse now.
Didn’t US defund some air traffic security thingy a few months ago?
FAA and the entire industry needs a proper modernization. $100 trillion was the said figure following the 2025 Potomac helicopter incident . Instead of funding or defunding, the current administration just attempted to employ ICE in place of TSA workers, other airport positions, and air marshals positions.
… billion dollars right?
I’ll attempt to find the source where I remembered that figure. The super high number was for yet-to-be-developed technologies that’s badly needed… (next generation radar, ground tracking, PCAS collision avoidance standardization, etc).
Instead of standard upgrades\market demands\expansions & construction to existing airports(1.8 trillion estimates)
https://www.businesswire.com/news/home/20250917931983/en/Airport-Construction-Strategic-Industry-Business-Report-2025-Market-to-Grow-by-%241.5-Trillion-by-2030-at-a-CAGR-of-14.2---Rising-Air-Travel-Government-Investments-and-Smart-Sustainable-Technologies---ResearchAndMarkets.com
This is actually a genuinely good idea since they’re not using LLMs to do it. It’s a very narrow problem that can be modeled and trained just like chess. Air traffic controllers can focus on the “shit’s hitting the fan” events
I knew the headline wasn’t the full story, but that’s better what I expected (LLMs with a deterministic verification method before making decisions to prevent hallucinations from causing plane crashes.)
Yeah “artificial intelligence” research was going on for at least half a century before LLMs, but the clueless will see this headline and think OMFG ChatGPT is GoNnA RuN tHe AiRpOrT!!1!
Air Traffic Control slop!!
(I said the thing, please applaud).
Yeah, it’s like people that thought Facebook was the Internet they now think ChatGPT is AI/ML. Here they actually can synthesise tricky situations where a pilot ignores instruction, comes in too fast and even if shit hits the fan automatically put in “circle around at x elevation while we focus on the emergency landing” command.
I read some time ago that air traffic control is pretty chaotic under the hood and having even just a minimal model that flags “problematic” sounds to me like a major upgrade.
But LLMs are AI/ML. They’re just one small example of it, though.
Sorry, I meant it the other way around in a “AI/ML is LLM” way.
It doesn’t help that 95% of these headlines do actually refer to LLMs. I’ve spent enough time on here repeating exactly that I hate the conflation of LLMs with AI in general (and the lack of acknowledgement that “AI” has never been well defined), and I still read that headline and thought “oh god no”.
Still hoping that the LLM bubble collapsing frees up a lot of both expertise and computational capacity to work on actually useful models like this one.
Since it’s all bunched up under the term AI the bubble burst will also be a drag on any other model research. It’s double edged.
I’m no ATC but I thought that to get good in crises, one has a train consistently…
Won’t this cut down too much on practice?
That’s a fair point and I don’t know enough about ATC to answer that. I’d imagine that if the job becomes easier it becomes also easier in a crisis but I have no clue if that’s actually true.
I reemeber when the Canadian military combined ATC and air defence. Both scope dopes controlling aircraft.
One keeps the dots apart and one brings the dots together.
Data might be corrupted.
I mean, this is potentially a good area where AI could augment human judgement.
Leave the human in control and have the AI note to them if they made an error or forgot about a plane or something.
Do not try and make it “self driving” or something.
My worry is that the hype was created by generative AI, so when they talk about AI I’m afraid they mean LLM.
Wouldn’t controlling traffic be a NP problem and we could have a hesitation algorithm to do it (ironically those algorithms we used to call AI).
I also would be surprised they already don’t have mechanisms to validate choices made by traffic controller. That probably be a polynomial difficulty problem.
I think the main reason why they want to introduce AI there is so they can fire traffic controllers as they are major headache for Republicans when they keep government hostage.
Luckily the people in charge seem to have great judgement with this sort of thing
You mean the A.I. in charge.
If you ever do stand-up, put me down for a ticket.
Depends on what level of “in charge” he means.
The career people who have run these things regardless of party in office have tended to be pretty good, historically.
Of course now many of these positions have been reclassified so that political stooges could be put in.
It’s unclear which level is actually implementing this and how much the Nero administration has their stupid, incompetent hands in it.
Give it few weeks, and they have collected meteics that the AI did the job just as good as the human within that timeframe. A few weeks down the line, the tables have turned, the humans are in charge of notifying the AI if it makes errors. A few weeks later, metrics will show that humans hasn’t really notified anything, and humans are removed from their position, leaving the AI fully in charge.
Within a year, a critical halucination or a job which AI lacks experience to handle will happen and a very preventable accident will occur. The owners will shrug it off as something that would definitely happen with humans in charge, and that the reduced costs outweigh the additional risk.
“Hallucinations” are produced by LLMs, which wouldn’t be appropriate for air traffic control. Automations of this type are done with Symbolic AI (aka “classical” or “logical” AI). Symbolic AI is built from sets of rules and conditions used by human beings to make the same decisions. This type of AI has been widely used in all kinds of control systems since the 1970s, and is completely unrelated to ChatGPT and other LLMs except by being called “artificial intelligence”.
AI, possibly. Algorithms, yes. Generative AI, fuck off all the way back to reddit where the training data came from.
High reliability organizations need deterministic tools that can be trained on so that operators can understand how they will work. They do not need an AI black box that might distract them with a hallucinated situation in the middle of a critical moment.
No one said this was generative AI, though. Which is good, because it’s not an LLM or genAI
AI/ML has long been used successfully to help predict future issues based on information from previous issues. It’s one of the things it excels at over humans, because it can process a much, much larger pool of data than a human can.
plus, AI is just a bunch of algorithms in a trench coat.
Then you have to say Machine Learning out front or everyone is going to assume you mean an LLM
It’s called Machine Learning. AI is synonymous with LLM crap now.
It’s actually called Symbolic AI, which codifies clearly defined rules and behaviors - think decision trees and flowcharts. I doubt machine learning will be used in this case, since the rules and reasoning used by air traffic controllers have evolved over almost a century and are well understood. Machine learning is more for situations where the software has to discover the best way to operate because the rules aren’t well established.
Machine learning is more for situations where the software has to discover the best way to operate because the rules aren’t well established.
Or the machine needs to look at messy input data and recognise what the situation is so it can apply the more structured decision processes. You can’t feed a radar image to a bunch of if/elif/elif/else statements. That pattern recognition is fallible, but so are humans. Exciting times!
You’d be surprised what symbolic AI can do with messy data - in medical diagnosis, for example - but LLMs suck at it.
Yes, because people don’t tune out when they have a machine doing most of the work. This is ridiculously stupid and dangerous. Don’t do work for them in trying to reason around it.
Although actual self driving cars are already better drivers than humans, although the bar is so low that Satan would have to dig for it.
Note, Teslas are not self driving. They’re mislabeled lane assist.
Where did you get that information from?
NHTSA reports human drivers to 1.9 incidents per million miles, and self driving cars companies refuse to give the mileage data. They are forced to give incidents rate data, and we can somehow estimate their milage and then we arrive at over 12 incidents per million miles for self driving levels 3-5.
Oh in 2025 NHTSA changed reporting requirements, to more relaxed, with less supervision and oversight, and no pushback or controls, laid off 25% of workers and changed it’s director to Morrison who was at the time Apple Self Driving Car Chief Council.
And even now it’s almost twice the rate of human driver per million miles.
That data includes Tesla
No, I excluded level 2, so excluding tesla shitty assistant
Note, Teslas are not self driving. They’re mislabeled lane assist.
What definition of “lane assist” includes the ability to manage city driving while following a gps route with turns and stoplights?
Better drivers than the average human driver. Take male drivers out of the equation and it tips. Women drivers may be less secure at times (compared to male overconfidence) but they are far less likely to be reckless or cause lethal accidents.
Signed: a male :p
No, it still doesn’t tip.
A computer can’t get intoxicated, tired, or distracted. Humans can’t monitor dozens of sensors millions of times a second.
There are definitely still steps that need to be made to advance the technology, but it already has us beat. And it’s great. Traffic related deaths are a huge issue.
What’s not great is who owns the cars.
Human drivers: people die due to reckless driving and human mistakes - and in most situations a good driver can anticipate a lot of dangers and avoid them. ML “assistants” getting control over cars: people die of unpredictable mis-assessments of inputs vs. outputs, and there’s no way to anticipate any danger or try to avoid it.
People who advocate for non-deterministic systems taking control in cars are irresponsible and have blood on their hands.
Your description of a human driver is no different than a self driving vehicle.
An input/output error is the equivalent to a human mistake. And those happen far less often than a human mistake. People make unpredictable stupid mistakes driving all the time.
If you think people make unpredictable mistakes all the time, you shouldn’t be driving a motorized vehicle. Because you clearly do not know what to anticipate. Do unpredictable mistakes happen? Sure thing. But they are the rare exception, and almost all the time an aware driver will be cautious about situations so as not to get caught unaware.
Just wait til the next government shutdown. All AI baby. They don’t need canned goods or toothpaste.
“AI” is such a terrible blanket term. Is this an LLM only? Will there be an actual machine learning model with hard checks doing the decision making? If the latter, I could see it largely working out. If the former … lol
Nope, it doesn’t even have an LLM component.
It just uses a ton of current and historical data to detect potential route and congestion issues.
This is something we have been doing for a while with the aid of computers. There has been a multi-year upgrade plan for this system in the works as well. Someone (Duffy) decided to take the already planned upgrade and just call it AI.
Definitely, there needs to be so much more info provided before anyone here could make a reasonable judgment
AI ≠ LLM
There are a bunch of other types of AI/ML that aren’t hallucination engines.
Yes but none of them should be let anywhere near decision making. When we talk driving assistants that actually work, those are driven by a hard set of rules, with deterministic decision making.
When they are trained for a specific task, they do noticably better. Hallucinations aren’t guaranteed to never happen but it is a lot closer to pure accuracy with niche AI. LLMs that you see are made for pleasurable conversation. AI/ML other than LLMs can become really to the T. I’ve seen YouTube videos testing such.
It’s like Ubuntu vs firmware. One is made to be able to do a ton of things, and one is made for a very specific purpose.
“noticably better” than what? You need to ask yourself that question. Machine learning solutions should never be in control of any irreversible actions. Slowing down a car due to perceived danger? Okay, worst case driver has to accelerate again (emergency braking decisions must not be made by such a system either to avoid being rear-ended). Swerving to avoid an obstacle? Absolutely not!
As I said in another comment: There are driving assistants that work deterministically. Anything else can give visual or auditory clues but shouldn’t be allowed to control much if anything else.
The question of how good an ML system is, is irrelevant as they are non-deterministic by nature, and their legitimate use cases are elsewhere.
Than humans and even expert humans. For specific, narrow tasks super human behavior has long been achieved in many classification tasks. Like tumor detection in XRays for example.
ATC jobs are extremely stressful and complex and they have surprisingly little digital support systems. A digital system could for instance easily suggest an optimal route through the airspace based on the aircraft’s capabilities and data and other airplanes in the air space before the pilots even checked in with the ATC. Today ATCs still have to request and suggest course changes via radio even though both them and pilots are sitting in front of computers.
You don’t understand the basics of safety rules.
Using bullshit language like “super human behavior” doesn’t make your point any less stupid.
Diagnostics are a valid application of machine learning - as a second check, after a qualified radiologist has reviewed the images. Not instead of the human, nor even as a first step, because it will impact the effort a human puts into a review if they already get a pre-analysis.
With your attitude, you will kill people.
Actually it’s safer when the human comes afterwards in the process to review and verify.
Please, stay dramatic, but I haven’t even mentioned replacement. If you look at my language I use “suggest” for example which means the judgement lies still with a human.
Actually it’s safer when the human comes afterwards in the process to review and verify.
So you really get off on the thought of getting people killed. Got ya, and e.o.t.
Well, at least comments in this thread can be excused because it doesn’t link to an article, just a stupid fucking reply.
WHAT IS ACTUALLY HAPPENING: AI will be used in the predictions about flight frequencies. IT WILL NOT BE IN THE TOWER.
(At least not yet. Which is what the articles are about. So at least realize that if you’re joking/speculating/bitching about ATC in the tower with AI, that’s not what this is)
I understand the HYPER ANTI-AI sentiments, but jesus christ people, try reading about what’s actually happening, yeah?
Thank you.
I know a good deal about machine learning. Machine learning’s principle metrics define how inaccurate it is.
The sheer volume of air plane traffic means you will have constant mistakes in those predictions. Leading to short staffing and actual consequences.
Staffing is not what they’ve said they’re using it for. Coordinating scheduling of flights to reduce waiting and conflicts is what I recall seeing.
They are extremely short staffed. Training takes 3 to 5 years. Very few eligible candidates and the Federal work force issues. This isn’t problem they can solve, because the problem is them, not the system.
I think we can all agree there are multiple aspects of a system as complex as aviation coordination in a large metro area that should be evaluated and improved upon when possible. I don’t understand how what you’re saying relates to what you’re replying to
Which has nothing to do with this.
AI being used by incompetent people trying micro mange their failure will not end well given the circumstance.
deleted by creator
deleted by creator
The streaming Mayday! Air Disaster channel is going to get its best comedic seasons.
“It’s not a crash. It’s an opportunity.”
“This was irresponsible behavior from me, and a betrayal of trust. I have adopted strict safety protocols to ensure that this cannot happen again”
. . .
< does it again >
Hooray for having a flight-heavy vacation in two weeks! It’s been real fellas.
Bon voyage, and, uh…bonne chance…
“Land all planes nose to nose, so every other plane lands backwards, otherwise it looks like a DEI.”
What
It’s not an LLM or GenAI if that’s what you were thinking.
