Hacker News Reader

Comment by aerophilic | Hacker News Reader

It is an interesting question: how do you plan for the unforeseen?

We as humans are very good at coming up with solutions to ambiguous/messy real world scenarios... robots less so. By definition, the current set of robots are programmed for specific situations.

As we automate more and more, I suspect we will have more and more “one off” catastrophic events. Yes, once we experience them, we can “test” against them. However it almost seems to argue for one amount of human in the loop no matter what, if only to give an occasional “that seems weird” check.

Am I missing something? Yes we can make an argument for machine learning to learn “nominal” and report off nominal... but that only works if you gave the right sensors being inputted in the first place.

View on HN

Fire in buildings and warehouses is not an unforeseeable risk. It is a major, well-known risk.

Correct, this particular case is obvious (and they should plan for it). However what I am trying to dig into is the “less obvious”. An example, snow showing up in an area that hasn’t seen snow in the last 100 years. Or perhaps someone “crashing a car” into your factory. My point, and my questions has to do with the following at the core: “How do we replicate/approximate the ‘general’ knowledge that a human will inherently have?”

Yep exactly.

I agree - humans are able to collate generally disparate, individually insignificant events to identify issues and unexpected events, but the entire idea behind ML is that AI is trained to react appropriately to the specific range of possibilities it is provided with. While ML allows for significantly better generalization within the trained problem set, it still flounders anywhere beyond predicted/tested parameters.

Hopefully, developing gaming AIs such as AlphaGo and AlphaStar will help us with this, but for now I remain dubious - AlphaStar, at least, has trouble reacting at all to the unexpected.

I agree current automation can't handle unforeseen events, but I think it can flag them as they start.

Where I work we use a hybrid system that must be pretty common — instrument, establish patterns, watch for anomalies (automatically) and alert humans when they happen.

We clearly just need fire smelling and fire fighting robots. And robots to repair and maintain said fire fighting robots.

So better smoke detectors and better sprinkler systems, in other words.

No, that's old tech. We need sprinkler robots!

we need sprinklers that leverage machine learning to adaptively choose different fire suppressants!

Simply use function as a catastrophe detector and create a generic quarantine and run plan?

Reply on news.ycombinator.com