4 ms·

Problem is first to adopt (assuming the AI is successful) will win. Imo, because most AIs are trained on real data, then augmented with simulations, im not sur

Problem is first to adopt (assuming the AI is successful) will win. Imo, because most AIs are trained on real data, then augmented with simulations, im not sure how well this will work in practice. Simple example, if it’s trained on historical data, it may try to use tanks where today tanks are not nearly as useful. Simulations would have to be VERY good and even then could have errors subtle enough for the Ai to fail. Finally, let’s say a country deploys this. It’s likely going to face nuclear annihilation. I really don’t see this working, not saying we couldn’t have a devastating effect short term (ie AI launches drones on battle field), but any sufficiently successful deployment would face destruction.

klipt 3y ago
> Problem is first to adopt (assuming the AI is successful) will win. Question is, will win what? Is potential human extinction worth winning?
- Filligree 3y ago
  What if someone else destroys the world first, though? What then? We'd better get a move on.
theptip 3y ago
How realistic are flight simulators? Not an expert but I had the vague impression that it’s pretty realistic in terms of physics, the main difference is the body feel (which isn’t relevant to an AI control system). In other words you could run simulated combat with pilots, and self-play vs other AI, and get a control system that works in most environments. (Maybe adverse weather and unexpected turbulence would be harder to simulate? Or maybe ot would be feasible to simulate crazy weather conditions that a human couldn’t train in.)
- meheleventyone 3y ago
  Realistic enough within normal flight but not at all at the margins and post-stall. Even a lot of CFD models used to assess designs aren’t accurate outside a small range.
  - theptip 3y ago
    So real pilots get good on the simulator then push the envelope & stall in training to learn those final edges? (Rinse and repeat?) Do you have a feel for what % of pilot training time is on the stick learning the real dynamics vs. simulator-based? (Excluding time flying that isn’t really used learning things an AI wouldn’t need to learn.)
    - meheleventyone 3y ago
      AFAIK almost all pilots primarily learn in real aircraft from the off rather than starting in a simulator and use of simulators is pretty limited.
      - trashtester 3y ago
        Hmm, I somehow feel that this may be true for people flying smaller civilian planes within their design envelope. But if training extreme manouvres outside the envelope for large passenger or combat jets seems both highly dangerous and extremely expensive. Like practicing to land a 747 with fires in two engines or to recover an F-35 from a flat spin.
        meheleventyone 3y ago
        That doesn’t contradict what I said. For example no one is going from zero to jumping in an F35 which is why there are several types of training aircraft military fast-jet pilots progress through.
vasco 3y ago
The DoD has been using Wargaming since WW2. There's many contracts already applying AI to this. Think of it as AI predicting how a conflict can go with millions of variables at play, doing it over and over and over again, kind of monte carlo'ing war and then immediately sending troop movement, attack recommendations etc to troops. But now imagine this instantly controlling drone strikes for drones that are already flying, to launching missiles from drone-submarines near targets, launching laser attacks from satellites or, you know, connect it to the wrong API and launching some nukes. The future.

klipt 3y ago

> Problem is first to adopt (assuming the AI is successful) will win. Question is, will win what? Is potential human extinction worth winning?

Filligree 3y ago
What if someone else destroys the world first, though? What then? We'd better get a move on.

theptip 3y ago

How realistic are flight simulators? Not an expert but I had the vague impression that it’s pretty realistic in terms of physics, the main difference is the body feel (which isn’t relevant to an AI control system). In other words you could run simulated combat with pilots, and self-play vs other AI, and get a control system that works in most environments. (Maybe adverse weather and unexpected turbulence would be harder to simulate? Or maybe ot would be feasible to simulate crazy weather conditions that a human couldn’t train in.)

meheleventyone 3y ago
Realistic enough within normal flight but not at all at the margins and post-stall. Even a lot of CFD models used to assess designs aren’t accurate outside a small range.
- theptip 3y ago
  So real pilots get good on the simulator then push the envelope & stall in training to learn those final edges? (Rinse and repeat?) Do you have a feel for what % of pilot training time is on the stick learning the real dynamics vs. simulator-based? (Excluding time flying that isn’t really used learning things an AI wouldn’t need to learn.)
  - meheleventyone 3y ago
    AFAIK almost all pilots primarily learn in real aircraft from the off rather than starting in a simulator and use of simulators is pretty limited.
    - trashtester 3y ago
      Hmm, I somehow feel that this may be true for people flying smaller civilian planes within their design envelope. But if training extreme manouvres outside the envelope for large passenger or combat jets seems both highly dangerous and extremely expensive. Like practicing to land a 747 with fires in two engines or to recover an F-35 from a flat spin.
      - meheleventyone 3y ago
        That doesn’t contradict what I said. For example no one is going from zero to jumping in an F35 which is why there are several types of training aircraft military fast-jet pilots progress through.

vasco 3y ago

The DoD has been using Wargaming since WW2. There's many contracts already applying AI to this. Think of it as AI predicting how a conflict can go with millions of variables at play, doing it over and over and over again, kind of monte carlo'ing war and then immediately sending troop movement, attack recommendations etc to troops. But now imagine this instantly controlling drone strikes for drones that are already flying, to launching missiles from drone-submarines near targets, launching laser attacks from satellites or, you know, connect it to the wrong API and launching some nukes. The future.