Give AI curiosity, and it will watch TV forever (2018) • Hacker News Board

Re: Give AI curiosity, and it will watch TV forever (2018)

#21

Post by samatman » Wed, Jan 31, 2024, 7:42 PM UTC

throwaway4aday wrote: post #13

Using prediction error as the definition of curiosity rings hollow for me. Curiosity in my mind is more about mapping out an unexplored thing and not about being surprised.

It's not a complete definition for a number of reasons, but it's a crucial component of curiousity.

If you're not surprised at any point in mapping out an unexplored thing, in what sense is it unexplored?

There are pretty high odds you've never been to this exact page before: https://oeis.org/A000079. But once you click on it, is there any remaining curiosity? It's an unexplored thing, in the sense that (I presume) you've never looked at this exact page before. But it doesn't provoke curiousity because there's nothing there to surprise you.

Re: Give AI curiosity, and it will watch TV forever (2018)

#22

Post by adammarples » Wed, Jan 31, 2024, 7:43 PM UTC

hahajk wrote: post #12

syntaxing wrote: post #5

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 w…

The researchers quoted were explicit about this: > 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 som…

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.

Re: Give AI curiosity, and it will watch TV forever (2018)

#23

Post by thatguysaguy » Wed, Jan 31, 2024, 7:45 PM UTC

thatguysaguy wrote: post #18

syntaxing wrote: post #5

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 w…

I think the problem is pretty interesting though. Better definitions of curiosity might still have this failure mode. Human curiosity definitely does!

As an example, look at the second link on HN right now: https://neal.fun/infinite-craft/

I just opened up a discord server I'm in and everyone is spending quite a lot of time on it!

Re: Give AI curiosity, and it will watch TV forever (2018)

#24

Post by gunshai » Wed, Jan 31, 2024, 7:45 PM UTC

thatguysaguy wrote: post #18

syntaxing wrote: post #5

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 w…

I think the problem is pretty interesting though. Better definitions of curiosity might still have this failure mode. Human curiosity definitely does!

Ya I think they are called loot boxes

Re: Give AI curiosity, and it will watch TV forever (2018)

#25

Post by p1mrx » Wed, Jan 31, 2024, 7:46 PM UTC

hahajk wrote: post #12

syntaxing wrote: post #5

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 w…

The researchers quoted were explicit about this: > 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 som…

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?

Re: Give AI curiosity, and it will watch TV forever (2018)

#26

Post by m3kw9 » Wed, Jan 31, 2024, 7:46 PM UTC

This is similar to how human minds work, things are bland when it’s predictable like stories, work like factories. We would tend to avoid it

Re: Give AI curiosity, and it will watch TV forever (2018)

#27

Post by photochemsyn » Wed, Jan 31, 2024, 7:46 PM UTC

Instead of TV, weather satellite feeds, and other streams of real-time constantly changing earth system-type data?

Also, stock market prices, but I imagine a whole lot of effort is already quietly going into that at present.

Re: Give AI curiosity, and it will watch TV forever (2018)

#28

Post by kemiller » Wed, Jan 31, 2024, 7:48 PM UTC

Can confirm this works for humans too.

Re: Give AI curiosity, and it will watch TV forever (2018)

#29

Post by jrockway » Wed, Jan 31, 2024, 7:49 PM UTC

gunshai wrote: post #24

thatguysaguy wrote: post #18

Earlier quoted context omitted.

I think the problem is pretty interesting though. Better definitions of curiosity might still have this failure mode. Human curiosity definitely does!

Ya I think they are called loot boxes

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.

Re: Give AI curiosity, and it will watch TV forever (2018)

#30

Post by pants2 » Wed, Jan 31, 2024, 7:52 PM UTC

howlin wrote: post #16

syntaxing wrote: post #5

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 w…

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…

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.