Hacker News Reader

Comment by londons_explore | Hacker News Reader

I'm not sure I agree here.

There exist possibilities for far more safety with more compute.

For example, when a mechanical failure occurs (for example an engine explodes and partially falls off) the flight characteristics of the plane change dramatically. Current systems fall back to human control, and hope that human will be able to figure out how to control a now very "unique" machine.

Future systems could dynamically create new flight models based on collected data to be able to fly the machine in entirely new ways.

For an example of this, did you know it's possible to fully control a quadcopter even with 3 out of 4 rotors broken and no working fins/flaps? [1]. That sort of 'fly it how it has never been done before' is out of reach of human control, and might save lives.

[1]:. https://www.ethz.ch/en/news-and-events/eth-news/news/2013/12...

View on HN

> For an example of this, did you know it's possible to fully control a quadcopter even with 3 out of 4 rotors broken and no working fins/flaps?

The linked article does not support your point, it shows a quadcopter losing one rotor which then goes into a kind of 'controlled fall', two (or even three) rotors is never mentioned and even the one rotor missing scenario involves the whole craft spinning (see video on the page).

This tech is just now making it into military aircraft in the F-35 (it’s designed to compensate for failures or combat damage while keeping the plane’s handling relatively consistent), so I would guess it’ll be a while before we see a civilian version.

"Figure out a way to fly the thing in the new, crippled configuration" is hard. Figure out a way that's better than what the human pilot came up with may be even harder.

It's hard, but it's a very measurable problem.

You randomly deform the airframe in a physics simulator, then get a real pilot to try to fly the simulated damaged airframe to the nearest airport.

You then get an AI to try doing the same.

As soon as the AI manages to get to the airport more often than the human, the tech will save lives, on average.

My guess is a PhD student could come up with an AI passing the above test inside a year. Yet aviation standards are stringent enough AI will be decades away from production use, if ever.

The political problem is that people still trust humans a lot more than they trust machines. So passengers would be nervous about being flown by an AI, even if the AI was safer.

There's some weird psychology behind this: human pilots/drivers/etc are seen by passengers as a personal proxy, with agency over any situation.

If you take away the proxy you take away the illusion of agency. Humans really do not like being put into situations where they believe they have no agency at all.

As AI gets smarter, this will become more and more of a problem, until there some kind of cultural shift because AI is obviously safer it's not even a question any more.

This may or may not happen.

Your link to ZTE trumpeting the success of an algorithm helping an actual physical model of a quadcopter with a disabled rotor operating in lab conditions to crashland a bit more softly might be interpreted as a hint that the general case of safely landing a stricken aircraft is the sort of problem that probably isn't solved by a single PhD in a year...

Keeping those with enough hubris to say "yeah, shouldn't take long if we've got some sort of sort of simulator to devise scenarios for a neural network to overfit to" about a wide range of failure modes that an entire field has spent decades studying well away from safety critical systems might be one of the more underappreciated aspects of tight regulation.

There also exists possibilities for far more danger. As you increase compute, you can run more code, and build more complicated interdependent components that are more difficult to test, especially for the sort of unexpected scenarios that "dynamic" behaviour would warrant. You would basically have a whole new plane - the opposite of what the 737 MAX 8 was intended to be.

Reply on news.ycombinator.com