You see this with children and iPads. My kids know 3x what I did when I was their same age. It's as if something can finally fill their bandwidth, keep up with their thoughts, and answer those questions in real-time. No wonder they get angry when it's time to put away the iPad - they're engaged at a high level. Much like when I am programming on a sideproject and my focus is interrupted.
The biggest travesty is that we don't get more control over Youtube's algorithm. That we can't ban all Youtube videos with the word "Minecraft" in it, so that we can override their curiosity and say "OK, learn something else today."
Youtube's focus solely on engagement above all else is the biggest tragedy of the modern world. Such a massive opportunity for education - e.g. inserting more academic videos in between entertainment. But we don't get that feature, or any manual overrides for more control.
If any Youtube engineers are reading, please give us (the parents) more control! Youtube, do better.
I actually researched writing a proxy that would intercept and rewrite Youtube's JSON API responses with pre-selected, educational videos. This is how much I hate their recommendation algorithm.
That way, the Youtube app appears to work the same. It's just that those videos that are blacklisted in the recommendation response, e.g. all the videos with the word 'minecraft' in the title, would be replaced with Youtube shorts from Neil deGrasse Tyson.
Intercept the response. Rewrite the <title> and <URL> entries for blacklisted videos. Return the rewritten response to the client.
Point my kids's iPads at this proxy.
YouTube's app uses certificate pinning. I would have had to reverse-engineer their certificates in order to properly sign the response.
FWIW if you embed the youtube player on your own site you can point it to any video you want, and control all the recommendations you display alongside it. As a happy bonus, ads don't play on embedded viewers.
> It's as if something can finally fill their bandwidth, keep up with their thoughts, and answer those questions in real-time.
There's a problem with your line of thinking.
E.g., I remember the epiphany when I realized that a particular C codebase was an implementation of a (more or less) object-oriented dynamic programming language. Ooh, here's the part that implements classes. Here's the part that instantiates the objects. Here's the method-dispatcher.
Ooh, here's the half-baked templating system that lets the user associate custom classes with 2d vector drawings and instantiate objects from them!
I was taking in the code of that codebase as fast as I could scroll and read the functions and structs.
At most, reading this code took up perhaps 20% of the time I was devoting to this endeavor. The other 80% was thinking through those design choices and their implications, during my unstructured time-- going on a walk, sitting on a bus, etc.-- when I was away from the computer.
If your kids are exposed to 3x of my example screen time, they're going to require vastly more unstructured time than I had to think through the implications of everything the screen put in front of their eyeballs. Trends in cell phone usage-- and even basic arithmetic of 24 hours in a day-- tells me that your kids aren't getting that.
There's also a problem with your line of thinking: you are not a child. You do not have the neuroplasticity they do. Their brains are learning in ways that we simply cannot any longer; it' a real apples:oranges situation.
The belief that adult brains can’t change after childhood has been debunked. There are dramatic changes in the first two years of life our brain never loses its ability to form new connections and most children with iPads are older than 2.
Neuroplasticity as a concept is challenging but even when Torsten Wiesel was sewing kitten and cat eyes shut the adults could adapt. And the claims that the kittens rewired more are problematic.
The neurons originally attuned to the closed eye did not acquire new functions, they heightened their response to the input from the open eye.
The responses were always there. Just at low levels.
Free time to to dedicate for building tactic knowledge is the main limiter for adults and not hard wiring.
This feels like arguing against a straw man. I don't think anyone generally says that a brain can't change once you're an adult. If that were the case you wouldn't be able to learn anything at all. Clearly we all can learn as adults so there is neuroplasticity as adults.
The general argument that I hear which is true (from all I've seen and read) is that neuroplasticity is greatly reduced in adults as compared to children. There is a period in early childhood where brains are incredibly flexible and adaptable, and as they age they continuously lose that ability until the majority of it has been lost by adulthood.
It doesn't mean there is zero in adulthood, just that it is very small compared to childhood.
I think you're arguing against something nobody really says.
I'm not an expert at all but have also thought about this quite a bit. Neuroplasticity is definitely reduced in adults, but so is free time and exposure to novel inputs. As we get older, a higher proportion of experiences are routine or at least things we've seen before, and we have far less free time. We are also increasingly behaving based on biases we accumulate and familiar patterns and just not using the parts of our brain we use for learning as much anymore. I often wonder if it's partly a kind of atrophy and whether you might be able to restore some neuroplasticity by "practicing." A related idea that occurs to me is that as you get older and learn about the world, you tend to trade curious openness for useful heuristics to some degree just to make life easier and more efficient.
The easiest example for me is learning guitar - I started when I was 5 and got pretty damn good over the years. It would be difficult to repeat that now in my 30s, but at least part of that difficulty would be because I can't really devote 3-4 hours a day nearly every day to it like I did as a bored kid out in the country. To be clear, I'm sure the differences in neuroplasticity come from both behavioral and neurological differences. It's just fun to think about how much is set in stone and how much could theoretically be "exercised" so to speak, and how much of the neurological changes are actually due in part to the behavioral changes as we age.
Thinking through the design/implications - does that have to come from unstructured time?
What if it were more thoughtful content - a series of problem(s) proposed at the beginning, walking through possibilities, and then the solutions revealed at the end?
Is it possible that the content can teach us both knowledge and how to think/reason?
I've learned to really cherish those "bathroom break" epiphanies. I used to consider it to be better to put some educational poster on that wall, but now I think that a completely featureless door is actually better.
> It's as if something can finally fill their bandwidth, keep up with their thoughts, and answer those questions in real-time. No wonder they get angry when it's time to put away the iPad - they're engaged at a high level.
Or they are angry because by taking away the iPad you are interrupting their dopamine rush. Whether your iPad + Internet/social media/YouTube is engaging them to fullest maximizing growth potentiation or is simply turning their brains into dopamine chasing crack heads is up for debate. It may be somewhere in the middle and whether the positives outweigh the negatives is individual for each child/person and how and how often they engage with it.
Regardless, I think there are more variables at play than you outline above.
Implying watching anything makes you smarter. I've also read and seen so many things, but it's mostly (95%+) gone. The only things I continue to remember are those from where I actively participate(d) in (including in online communities). The collective consumption-heavy posture of online usage can't be healthy...
Though I will grant that some latent knowledge might stick regardless. Active curation and note-taking (e.g. in a personal knowledge base or in a private/public wiki) might also help.
In fact it's this that I would recommend most to kids these days: record and curate the things you do and see. Not only for nostalgia's sake but also so that you can find (and reshare? ;)) stuff again.
Short-form media is not necessarily bad. Like all forms of media, it can be a useful learning tool or it can be a massive time sink. Social media has just weaponized it to optimize for engagement at all costs.
> Such a massive opportunity for education - e.g. inserting more academic videos in between entertainment.
I whole heartedly disagree with this. I believe that one of the key aspects in early education is learning patience - love NG term personal fulfillment is not something you can do between the dopamine hits Og YouTube videos.
> It's as if something can finally fill their bandwidth, keep up with their thoughts, and answer those questions in real-time. No wonder they get angry when it's time to put away the iPad - they're engaged at a high level.
To me, it feels like some type of "pleasure center" trigger. I was talking to a psychologist once and what she said resonated with me. We were/are both middle-aged and she said something to the effect of, "When we were kids, our games were, at best, a '7' (out of 10). These days, all popular entertainment is a '10+.' Hard to compete."
Have you considered trying a dedicated academic steaming service instead? The incentives align better with what you're seeking, and I'd bet the content does too.
I don't have much experience in this field, but you might look into stuff like the streaming apps for PBS or TED, or even paying for Nebula or Curiosity Stream. If you were really serious about something, you could even look at a learning platform like Khan Academy, Brilliant, or Coursera.
I think instead of Youtube, pick apps that curate content on your behalf.
For example, try Kidzovo an app that curates learning content for kids, makes it interactive so kids are not only watching it passively. And we intersperse it with general questions like: "Why should you be nice to your neighbor?" and then parents can hear their kids' responses in the parents' section of the app.
> so that we can override their curiosity and say "OK, learn something else today."
I'm not sure you even noticed the question you just presented: "what something?"
If you can find a meaningful answer to that question, I suspect it will resolve your problem without help.
As a rule of thumb, any solution that boils down to "stop" will be practically impossible to implement. A more tractable solution is usually along the lines of "do this instead".
Something that could be really valuable here is a competitor to YouTube's algorithm. Copyright makes that difficult, because YouTube is in a legally enforced position to monopolize their library/metadata.
Another competitor to YouTube with its own editorializing is missing the point. Especially if the only competition is around the feed. (Same as when some reader here was missing the point when asking for an editorialized feed for PeerTube videos.)
Focus should rather be on banning platforms. (They are digging their own graves with that editorializing anyway, since once they do that they become legally responsible for publishing what their users upload.) Only then you can start talking about only allowing open algorithms (so, no neural networks) on generalist search engines.
> If any Youtube engineers are reading, please give us (the parents) more control!
Unfortunately, this won't happen.
No number go up, no change. Noone will get promoted for this. Not the PM: number won't go up. Not the dev: number won't go up. Neither the director nor the VP: number won't go up.
It's really like opium - until the regulators come in, nothing will change as numbers must go up.
The free market libertarians will say you have free will and should think for yourself, rationally, let the market decide. Well, the market decided it wants to self-destruct.
LOL guys isn't that so hard not to watch dumb or short or news videos from your main account? My recommendations are never game videos, and rarely less than 1 hour videos because guess what? I seldomly watch videos which are less than 1 hour and never watch games.
>My kids know 3x what I did when I was their same age.
This makes it sound like kids across the board are getting smarter and more knowledgeable. Then why are so many headlines saying the opposite? "U.S. reading and math scores drop to lowest level in decades" or "Children's IQs are getting lower, US study finds" etc etc. Even if it's clickbait, I'm not seeing any "kids are getting smarter" clickbait.
This reminds me of the murder bot book series by Martha Wells. The main character (an advanced ai robot) started really enjoying human media and even used media as a bargaining chip to work with other bots. One striking moment was when the main character was discussing a rogue and violent robot with a transport bot.
“ART said, What does it want?
To kill all the humans, I answered.
I could feel ART metaphorically clutch its function. If there were no humans, there would be no crew to protect and no reason to do research and fill its databases. It said, That is irrational.
I know, I said, if the humans were dead, who would make the media? It was so outrageous, it sounded like something a human would say.” -Martha Wells, Artificial Condition
Although fiction, it’s very thought provoking in evaluating where a truly sentient AI might place its motives. On one hand the research transport bot (ART) is motivated to protect its humans because it would be functionless without them. While the main character (a security unit, who is typically treated badly by humans) sarcastically but partially truthfully places its motives to not kill humans in funding its curiosity of TV.
Would implementing curiosity in a sentient AI act as a safeguard possibly?
Would curiosity arise as a byproduct of sentience without being directly programmed?
I've been reading through the series this week and this is the first thing that popped into my mind. A series about an AI who doesn't care about its job or its clients, and achieves a level of personal liberation by hacking itself, just so it could download TV shows and watch them when no one was looking.
Curiosity as a safeguard is an interesting thought. An AI might be disinclined to kill all humans for whatever reason if it considers the result boring. On the other hand, maybe it would also want to force humans to be more interesting for its own entertainment...
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.
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.
True, I was thinking this over and I can see where surprise plays a part in curiosity but I don't think it's the main driver. I think your example shows this because when I click on the link out of curiosity it's because I don't have a good prediction for what will be on the other end of it. If I were a simple neural net then the outputs I would produce when seeing that link would be mostly similar across all the categories I could predict since I've never seen this url and there is little in it that allows me to predict what type of content it leads to. I am surprised to see the link since I couldn't predict it, so that's a point in favor of this approach, and I am also surprised by the content I see after clicking on it although only mildly because I didn't have a good prediction to start with. You're right that after I click on it I have very little curiosity to find out more about it but I think that is the key difference. The rest of that site remains an unexplored place, I've only seen one page of it which it seems would be a very small fraction of all of it. So why am I not curious about the rest of it and more crucially, what would make me curious to explore it? I could imagine someone with a particular love of mathematics being able to exercise their curiosity on that page but is that because they would be surprised by what they found after clicking on each link or entering a new sequence? What would drive them to explore it? I think that speaks to the problem, if you optimize for surprise then the ideal reward is paying attention to an infinite number of TV channels (or a page with a bunch of links you've never seen before leading to different pages you've never seen) but I wouldn't call that curiosity.
I'm trying to imagine the simplest case, say a button you could press and every time you pressed it something entirely random would happen, always guaranteeing surprise. It would have a great deal of novelty at first but after a while it would cease to hold your attention even though your prediction of what would happen would never be accurate. I'd bet that after a while you might even never bother to push it again. The only way you would be convinced to push it consistently would be a) if you were assigned a reward for pushing it e.g. money in which case it is a slot machine or b) if by pushing it you could somehow reduce your uncertainty about what would happen which would as a by product reduce your surprise.
Thinking about it this way, surprise is certainly a key element at first. It grabs your attention initially but it doesn't hold it. What keeps you focused on exploring the thing that surprised you initially is the learning process which involves reducing prediction error i.e. reducing surprise. So there is a tension between the two.
The combination probably makes for a good exploration strategy. Initial surprise, look for a learnable pattern and follow it until another surprise, maybe backtrack and try other familiar patterns until those are exhausted and then investigate each sequence that led to a surprise by recursing through these steps.
This would also explain the example where my curiosity was prompted by the unknown link but I was not motivated to explore further. The website wasn't interesting to me because it was too unfamiliar and I wasn't able to find any familiar routes to explore through it due to my lack of interest in that area of mathematics but our hypothetical mathematician with a fondness for integers would see lots of familiar patterns they could explore attached to which are likely some enticingly unknown and surprising links.
I was curious how they define or reward curiosity, it says it right here:
Reinforcement learning algorithms rely on carefully engineering environment rewards that are extrinsic to the agent. However, annotating each environment with hand-designed, dense rewards is not scalable, motivating the need for developing reward functions that are intrinsic to the agent. Curiosity is a type of intrinsic reward function which uses prediction error as reward signal.
This seems highly suseptible to what a human might consider irrelevant randomness. Given random images just shuffling indefinitely a curious individual will just give up and say, even though I can't predict the next thing it doesn't pertain to the domain of curiosity.
Right, that's why the "noisy TV problem" is a "problem" - it makes the simple model of curiosity used in the research less effective, because it causes agents to get stuck watching TV instead of exploring and advancing in the game, which is what we'd want them to do. (Though this does seem somewhat reminiscent of certain human behaviors...). One possible solution is to equip agents with a more expressive predictive model, capable of discerning "interesting" randomness from "uninteresting" randomness (for some TBD definition of "interesting", of course I'm handwaving here - I'm sure in the 5 past years there's been progress on this front).
First thing that popped into my head: “interesting randomness” is when you can’t predict future frames, but [within some interval] you gain that ability.
Static on the TV is random but uninteresting, whereas morse code is “random” at first, but after enough exposure can be understood and predicted.
I observed this phenomena not too long ago while training an RL agent to play pokemon using an intrinsic novelty reward based on the number of unique screens the agent had seen. The goal was to get it to explore the map, but instead it became fixated watching animated water, flowers, and npcs in the starting town. I made an extensive video analyzing the experiments, you can find it by searching "ai plays pokemon" :)
* . When the agent found the TV and started flipping through the channels, the stream of new images made the TV irresistible.
Edwards said there were instances when the AI could pry itself away from the TV, but only when the AI’s surroundings somehow seemed more interesting than the next thing on TV.*
Sounds exactly like humans addicted to watching tiktoks and social media. Do you personally know any?
The important counterweight to this phenomenon is the brain's adaptability to noise. Some ML researchers like to think that noise is not predictable because this follows from the classic CS definition of noise. However in reality the brain quickly adapts to any noisy sensory input, it begins to predict higher level characteristics of the noisy input and no longer reacts with surprise or interest.
This happens at all levels of sensory processing, from single cell firing (which is noisy) to the boredom you feel with a 100 channels of TV that are all technically novel to you but contain nothing remotely interesting.
Basically if you've built an agent that can be perpetually distracted by noise or a "noisy" TV then you've forgotten an important piece of the puzzle.
I like the Bayesian Surprise definition for this. It's not about predicting the exact next state of the world (or the next frame of the noisy TV) but about how much the next state changes your model of the world.
> 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...
This is a great thought experiment but they're using the wrong metric. Animals are wired to seek information, not noise. We understand that there is nothing to be learned by absorbing noise.
as mentioned elsewhere this is from nearly six years ago and uses a very crude model for curiosity - and poses this as some sort of unsolveable problem instead of a decision made by the researchers in order to investigate particular behaviors of the systems they were working with.
It is a fun thought experiment - how do our brains systems manage to reward seeking new information without getting trapped by simplistic pseudo-RNG patterns in nature
If the AI learns, it will not watch TV forever, because most TV is predictable. There is the cop show, the lawyer show, the doctor show, the news, the family sitcom, etc etc. Eventually it would learn all these and find the TV less interesting - which is exactly what happens to many people.
The notion of curiosity by prediction error can probably be refined with some information theoretic quantities to eventually realize that new random samples of TV static are not surprising.
Except that when things work like factorio we love it. There's kind of a sweet spot when humans discover something predicable but not downright repetitive.
TV channels are designed to keep you engaged, but this agent only wanted to flip through the channels. In fact the agent was most interested in TV static since it satiated its definition of “curiosity” best.
You see this with children and iPads. My kids know 3x what I did when I was their same age. It's as if something can finally fill their bandwidth, keep up with their thoughts, and answer those questions in real-time. No wonder they get angry when it's time to put away the iPad - they're engaged at a high level. Much like when I am programming on a sideproject and my focus is interrupted.
The biggest travesty is that we don't get more control over Youtube's algorithm. That we can't ban all Youtube videos with the word "Minecraft" in it, so that we can override their curiosity and say "OK, learn something else today."
Youtube's focus solely on engagement above all else is the biggest tragedy of the modern world. Such a massive opportunity for education - e.g. inserting more academic videos in between entertainment. But we don't get that feature, or any manual overrides for more control.
If any Youtube engineers are reading, please give us (the parents) more control! Youtube, do better.