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
It’s like how there are models that are trained principally on medical imaging and diagnoses. They’re getting very good at detecting disease. It’s what AI should be doing, not infecting as much user-facing software as possible, with shit that doesn’t have a net benefit.
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!
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.
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.
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.
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
It’s like how there are models that are trained principally on medical imaging and diagnoses. They’re getting very good at detecting disease. It’s what AI should be doing, not infecting as much user-facing software as possible, with shit that doesn’t have a net benefit.
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.
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.
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.
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.