[mlinhares]

Such an incredible write up, the piece about the importance of flying less technological planes to get a "sense" of what flying really is hits like a brick, specially in the world of LLMs producing code.

How do you get this "sense" of writing code and building systems by yourself if all you do is instruct some agent to do it? Are we all going to be like Bonin in the future where we just don't understand anything outside of the agent box?

This is both terrifying and sad.

[ottobonn]

I'm a software engineer and recently got my pilot's license, and the training for the pilot's license increased my (already-high) respect for the aviation profession. All pilots learn to fly basic airplanes and have to do everything by hand (often on paper, but an iPad is allowed) to show they know the basics. The result is that by the time you work up to more advanced planes you have climbed the ladder of abstraction and know what underpins the automation.

The other piece of the picture is that pilots acknowledge that their skills are perishable, and they have to commit to ongoing training. This would be analogous to writing code by hand and getting a licensed engineer to sign off on your currency periodically even if you use LLMs for work.

[altmanaltman]

But I mean flying a cessna vrs something that has fly-by-wire like Airbus jets, its not really about understanding abstractions or anything, since the plane is basically a fundamentally different machine no? Basic principles of gravity and physic apply sure, but the flying experience is 100% different and not like a levelling up thing right? Like i would not trust someone with a Cessna pilot license to fly the airbus i am on.

[OneMorePerson]

I'm not a pilot (someone correct me if I'm wrong) and I respect how hard it is to make split second high stakes decisions, however from my read the pilots had to ignore tons of basic sensors (including their own bodies). According to the crash report they were at a 35 degree upward tilt, which is super severe, and they thought they wouldn't crash because they had successfully recovered from a "low altitude warning" (which doesn't makes sense).

Point being it reads like following the sensors but ignoring what you can actually feel is happening, which is back to fundamentals.

I don't think people are saying a Cessna translates to flying an Airbus, more that NOT knowing the basics or forgetting them translates to dangerous gaps when flying the Airbus.

[wolvoleo]

Well, they were in IFR. It was dark and they probably couldn't see the horizon. The senses can play tricks on the body when you have no frame of reference. This phenomenon is trained for of course but it is hard to avoid.

However what I do think they should have realised is that whatever they were doing (pulling up) did not work and maybe they should stop for a moment and think about their assumptions. It's in fact hard to understand what situation the plane could be in for a hard constant pull up to be the right answer. The only thing I can think of is a loss of vertical stabiliser trim, a bit akin to what happened to that Alaska airlines crash off the coast of LA. Or a sudden extreme shift of cargo forward. But then that assumption could be checked.

But the mind can also get into a state of panic that makes such reasoning very difficult. That also is being trained for. But it is still very hard to overcome.

[WalterBright]

A Cessna has very different aerodynamic issues than a jetliner. Multi-engine also has its own issues (such as if one engine dies, the airplane tries to turn around it).

Setting a Cessna down on the runway is fairly strait forward. A jetliner, on the other hand, is quite complex to land.

[VBprogrammer]

I don't know if you can claim one is more straightforward. Sure a Cessna flies slower and has relatively simple aerodynamics. However, you could also be operating it out of a 400m sloping grass strip with a mountain off one end.

An A320 might be flying 3 times faster but is generally flying between relatively flat, straight runaways several miles long with approaches typically flown on a stable instrument approach from several nautical miles away. It's control laws mean flying straight or maintaining a particular bank is as simple as letting go of the control stick. If anything the stick and rudder skills in normal circumstances are much less involved. Systems management, obviously the autopilot, but also environmental, hydraulic, navigation an the operational concerns are obviously vastly more complex.

[defrost]

> you could also be operating it out of a 400m sloping grass strip with a mountain off one end.

Why? Not as a regular thing I hope, that's about 90m short of "tight".

If you're intent on proselytizing PNG at least get a PAC STOL ( https://en.wikipedia.org/wiki/PAC_P-750_XSTOL )

[VBprogrammer]

Not if the Cessna is a King Katmai 182...that would have room to spare.

[raverbashing]

I see where you're going here but no

A Cessna an a big jet fly by the exact same principles and they stop flying due to the exact same principles as well

Sure the procedures and parameters and automations are different (as well as things like wing positioning, engine positioning, swept wings, number of engines, sure)

But you raise the nose of both of them enough they will both stall. If you lose speed they will both stall. They will behave similarly (or maybe weirdly) enough in curves.

And I think this is what was forgotten here. Having a fancy cockpit does not make it less than a dual-engine swept-wing fixed-wing aircraft. The principles are the same

[WalterBright]

Flying at near supersonic speeds at high altitude with a swept wing airplane is quite different from low and slow with a straight wing and thick air. Jetliners have a rather small envelope at altitude where the airplane will fly, things like overspeeding it will cause it to go out of control.

A fair amount of effort goes into designing the cockpit so it feels to the pilot like a low and slow aircraft, but it is not the reality.

For example, jetliners are unstable and require a yaw damper.

[raverbashing]

Yes yes and I know all of that

We can argue semantics but the reason AF stalled is the same reason a Cessna would stall (too high of an AoA)

And fair enough the comparison with the Cessna might be bad, but compare it with a 737-200 or even an A300 and the comparison will be much closer even though the 737 doesn't have the fancy cockpit

> things like overspeeding it will cause it to go out of control.

Well that would be the case of the Cessna as well, if it has enough power

[thrownthatway]

I’ve flown a couple single engine aircraft.

I put it this way:

Commercial aviation pilots don’t really fly the plane as such. It’s more like a 1:1 real-time flight sim. They’re sort of up there having a LARP.

They’re flying in a similar sense that a DJ creates music.

[smcin]

But it wasn't at all just about Bonin: Robert and Bonin repeatedly kept trying to override each other; Robert was giving Bonin some information with which he could have figured out he was stalling (although Robert was also trying to climb); and Dubois had gone to deal with his sleep deprivation without designating either of them as PIC, and when Dubois finally returned (2:11:40, sounds still sleep-deprived based on his confusion) he didn't recognize the obvious stall or take control until they had lost almost all of their altitude.

It makes Air France look worse that all three pilots didn't react properly (or weren't trained or experienced), than just faulting Bonin. And at least two of them were sleep-deprived. But there were multiple systemic failures, not just the pilots.

[OneMorePerson]

I think this was a pre-LLM issue, but will be exacerbated. I've worked with some people who would be reading logging data and about to do some big intervention when it was clear that the logger itself was having an issue and the results coming out were not really possible.

[altmanaltman]

The irony of not understanding almost 100% of the code on modern airplanes is actually done by instructing a program to actually generate the code. It is neither terrifying nor sad. You expect humans to write millions of lines of code? At that scale, procedureally generating code is much safer and smarter.

[flohofwoe]

Those millions of lines of code can often be reduced 10x or 100x with just a bit of common sense, and with that also reducing the potential bug count by 10x to 100x.

Also unlike LLMs, traditional code generation techniques are deterministic.

[altmanaltman]

You think simulink code is not deterministic? And okay tell me how will you optimize an Airbus 320's software down to 10x or 100x? I'm waiting.

[hunterpayne]

I'm not flying anymore if that's the case.

[deepsun]

Novella "Profession" by Isaac Asimov.

[riffraff]

"Profession" is often cited with regard to LLMs, but honestly, in reminded more of (and scared by) "The Feeling of Power".

[cladopa]

Actually there are more planes flying today than ever and the number of accidents is very very low, thanks to technological planes and protocols that lean from mistakes.

So low in fact that the majority of the recent "accidents" look like suicides from the pilots. The pilots know exactly what they are doing when crashing the planes.

[thrownthatway]

Boooo!