4 ms·

I agree with your first point, such objective function would have to optimise something extremely abstract like player enjoyment. As for the second, the point

by qqii 7y ago

I agree with your first point, such objective function would have to optimise something extremely abstract like player enjoyment. As for the second, the point is not to have a specificition on the human side. Most of the time when communicating with others we don't need lawyers, and spoken contracts are valid and by law. Even with lawyers contracts can be disputed in court as there is not a formal definition of every exact scenario. Assuming we had the power to do (1), all we'd need in (2) is something that doesn't provide unexpected outcomes.

mattkrause 7y ago
If you don't have some kind of specification, how are you going to control what you get? It's true that you can turn to a teammate and say "hey, can you write me a pipeline to import data from JSON files?" and you'll usually get something usable. However, you and your teammates have shared goals and background information about your particular project and the world at large, etc. Projects go off the rails all the time because the generally intelligent (allegedly, anyway) humans don't share these things. Right now, the front page has an article called "Offshoring roulette" about this. If you don't want to click through, here's my story about a contract programmer working in the lab. I asked him to look into a problem: the software running our experiments crashed after a certain number of events occurred. It turned out that the events were being written into a fixed-size buffer, which was overflowing. This could have been fixed in many ways (flush it to disk more often, record events in groups, resize the buffer). However, he chose to make the entire saving function into a no-op. This quickly "fixed" the problem--but imagine my delight when the next few runs contained no data whatsoever. In retrospect, although this guy had a PhD, he wasn't particularly interested in the broader context, namely that it crashed while collecting data that I wanted. An optimizer is going to take all kinds of crazy shortcuts like that unless it's somehow constrained by the spec. You could certainly imagine building lots of "do-what-I-mean" constraints into this optimizer but that requires even more magic.

by qqii 7y ago

mattkrause 7y ago

If you don't have some kind of specification, how are you going to control what you get? It's true that you can turn to a teammate and say "hey, can you write me a pipeline to import data from JSON files?" and you'll usually get something usable. However, you and your teammates have shared goals and background information about your particular project and the world at large, etc. Projects go off the rails all the time because the generally intelligent (allegedly, anyway) humans don't share these things. Right now, the front page has an article called "Offshoring roulette" about this. If you don't want to click through, here's my story about a contract programmer working in the lab. I asked him to look into a problem: the software running our experiments crashed after a certain number of events occurred. It turned out that the events were being written into a fixed-size buffer, which was overflowing. This could have been fixed in many ways (flush it to disk more often, record events in groups, resize the buffer). However, he chose to make the entire saving function into a no-op. This quickly "fixed" the problem--but imagine my delight when the next few runs contained no data whatsoever. In retrospect, although this guy had a PhD, he wasn't particularly interested in the broader context, namely that it crashed while collecting data that I wanted. An optimizer is going to take all kinds of crazy shortcuts like that unless it's somehow constrained by the spec. You could certainly imagine building lots of "do-what-I-mean" constraints into this optimizer but that requires even more magic.