With the definition of curiosity from the article, it’s not that surprising? A dynamic “screen” is always more interesting than the static map.
Definition: The definition that OpenAI team used for artificial curiosity was relatively simple: The algorithm would try to predict what its environment would look like one frame into the future. When that next frame happened, the algorithm would be rewarded by how wrong it was. The idea is that if the algorithm could predict what would happen in the environment, it had seen it before.
There's clearly a sweet spot in the amount of entropy/unpredictability that is "interesting". Otherwise observing white noise would be the most interesting thing imaginable.
I don't know the details, but probably you would want to seek unpredictability in a higher level representation of the observed state. White noise is highly unpredictable per pixel, but will get a very predictable representation after a layer or two of featurization if the features are trained/designed for real world observations.
I think there's a gap between the human version of curiosity and the AI version. A machine can be told that something is interesting, where humans need to innately find something interesting or spend a long time sort of learning to find something interesting.
> White noise is highly unpredictable per pixel, but will get a very predictable representation after a layer or two of featurization if the features are trained/designed for real world observations.
Virtually anything that cannot be predicted is interesting by nature of being unpredictable. Is it truly random? How, or why? True randomness is rare, and its existence is interesting.
TV static is uninteresting because it isn't actually random, it's just too onerous to get the measurements to predict it for the value we would get. It's part of the large class of things that is random for practical purposes, but not truly random. I have no doubt that if humanity dumped all its resources into predicting static, NASA could measure inbound radio waves and/or model space to figure out what static would look like at a particular spot.
Notably, humans find the cause of static (partially various waves from space) fascinating because we can't predict them. We've just placed our interest down a layer of abstraction from static. Static is boring, the source of static is interesting.
I suspect it is truly random to the AI, though, because it has no means to "see" those radio waves. I would wager humans would be far more interested in static if we were also unable to see the causality between radio waves and static.
I would be interested to see if the AI was as interested in static if it was also provided a real-time feed of radio waves at the antenna. Would it figure out that those things are correlated and lose interest in static like humans have, or would it continue to find static fascinating despite knowing it's a basic causality?
> A machine can be told that something is interesting, where humans need to innately find something interesting or spend a long time sort of learning to find something interesting.
Humans seem to be the same way. Lots of people learn something because it pays well.
It's possible that white noise is interesting to look at but it simply overloads our feeble human brains. If you could zoom in, slow down, and blur the white noise to make it a slowly changing gradient I bet it would be somewhat engaging.
> OpenAI researcher Harri Edwards tells Quartz that the idea for letting the AI agent flip through channels came from a thought experiment called the noisy-TV problem. The static on a TV is immensely random, so a curious AI agent could never truly predict what would happen next, and get drawn into watching the TV forever. In the real world, you could think of it as something completely random, like the way light shimmers off a waterfall.
The headline is really just inappropriate anthropomorphization.
The article seems pretty incoherent. It's not clear if the AI was watching static or actual TV content. If it's static, then why bother flipping through channels?
Perhaps it should use prediction error on some higher level of embedding, that way boring changes like static would be treated similarly but genuinely novel things would be treated higher.
To me the real question is if humans are really much more complicated. We evolved running around on the plains without TV or drugs or electric guitars or virtual worlds. How long until we completely crack our definition?
That's not curiosity, that's trying to get a reward through randomness. Loot boxes are like being hungry, going to your kitchen, and picking 3 random ingredients to combine. The mayonnaise, raw onion, and ice cube soup is not so good. So you try again. Eventually you land on cooked spaghetti, butter, and cheese. This encourages you to keep trying.
Curiosity is more like scrolling on social media. You know there have been interesting things there before, so you keep looking for more interesting things.
Definition: The definition that OpenAI team used for artificial curiosity was relatively simple: The algorithm would try to predict what its environment would look like one frame into the future. When that next frame happened, the algorithm would be rewarded by how wrong it was. The idea is that if the algorithm could predict what would happen in the environment, it had seen it before.