Pretty much the most succinct way to make the point that there are really direct ways to influence. Nice linguistic twist. BTW, did you intentionally misspell the brand? Gotta think you did.
Of course this is directed to someone who already has a habit, but the same techniques apply to recruiting. Aim Hi.
Good piece, but I can't help feel as though you've built a bit of a false dichotomy here.
I fully agree that it's ridiculous for the Facebooks of the world to claim they've "solved the dopamine system" or something of the like, but the idea that the networks aren't shaping their users' minds (on the margin) seems so obvious to me that I almost feel like the burden of proof is on someone claiming that they don't do this.
That said, I feel like the breakdown comes between the difference between being able to control users' minds in the abstract vs being able to influence them on certain ideas. I do not think the Zuck army can influence me to be something I'm not, but I'd be surprised if they didn't have the ability to push me in certain directions on certain axes.
As example, my dad is a Fox News diehard - spending time at his house means hearing Tucker et al at all hours of the day, and I find that spending 2-3 days around there starts to do interesting things to me. I don't believe the nonsense they often spew, but I can sense myself becoming gradually sympathetic to certain ideas, or agreeing with certain framings of things.
Abstracting this to more subtle changes in ideology and longer periods of exposure (as I'm sure you know, people use FB a _lot_), it's hard for me to believe that people can't be pushed. Whether or not this is being done, or whether it's profitable or anything remains an open question to me, but IMO the answer of "could TikTok influence certain thoughts (I imagine all of us are more suggestible on certain topics) of someone who watches it for 3 hours a day" is a resounding yes.
Thinking of things this way, I guess my condensed response would be something like "ABC/NBC/CBS could certainly claim to influence the zeitgeist (or could've 20 years ago) but could not influence me to like them more than the internet because that's out of their axis of influencability", which I think is maybe somewhat congruous with what you're saying but not entirely.
I think so! I think Iβd insist that theyβre not capable of even intentionally influencing things, but Iβd also accept if people didnβt believe this.
I remember FB did some experiment where they attempted sentiment analysis based ranking and then claimed that additional sentiment analysis of posts from those exposed to sentiment ranking analysis showed that they were influenced. My own read of this experiment was that it showed nothing: sentiment analysis is unreliable in both cases βfor example, it cannot make sense of sarcasm lmfao, or couldnβt then (but probably could now!).
But if I were a betting man, Iβd bet close to your POV here; Iβm sure Iβm overstating things, probably in response to overstated things!!!
That's a great point! I think you can look at the causation al kinds of ways; I've had Ben's experience of feeling influenced many times, and I wouldn't dispute that this is a common phenomenon.
But I also think if you asked most users of these products / services, they'd say that they they came to them because they matched beliefs they already had The path the critique posits is: a given person is exposed to these services, then captured by them, then has their beliefs or will shaped by them. But I think another perhaps more common (and perhaps in many cases simultaneous!) path is: a given person knows these services are where people with the beliefs or habits they have congregate and learn the "best" interpretations of new events according to those beliefs / stay up to speed with their community.
For example: in the case of Fox News, a deep belief many conservatives have is something like "there's a group called liberals who hate me and my beliefs and way of life and want to change everything about my world and I want to oppose them at every step." People with this value (and similar ones) watch Fox News because Fox News is what people with this value watch and where their worldview is substantiated and applied; the specific claims Fox makes, which they then adopt, are not actually that important to them, which is why they can often drop them like a hot potato when it's expedient for their war with liberals.
We imagine when a news story breaks that without Fox News, these people would have a different take than the one Fox News promulgates, but I think the causation runs the other way, as the Dominion lawsuit shows: Fox News tries to intuit what its audience wants it to represent, while much of its audience wants it to represent what's most useful for its psychological purposes + broader war against their enemies (which is something that they in part determine by... watching Fox News and seeing what e.g. Sean Hannity thinks will hurt liberals).
A lot of my skepticism about censorship comes from how I perceive the causation here. I can't deny that influence happens, of course, but I do deny that it is stable, deterministic, highly or reliably deliberate, or unbounded.
Yes exactly. The company may not be the βintentional initial influencer,β but it might be the βunintentional secondary spreader.β
For instance: if 12 guys used to go into the woods and be a hate group, now they can go on Twitter and find all the other hate groups and be like βyeah we all hate the same thing!β Before they didnβt have power and now they do. These platforms may not be intentionally creating that, they are just unintentionally empowering that?
(And I would imagine they are doing so for good things just as easily as for bad things. Grassroots movements being able to find each other etc......)
Brilliant Mills. Still, we're pretty good at hacking ourselves. A progression might be email, chat/dm, social media, tiktok. The parallel universe is (yeah mentioned this before) right wing mass media, also a hack. Judging from our kids, it doesn't take all that much work to write an algorithm to keep us scrolling for hours. The length matters. Short catchy content seems to command a lot of eyeballs. Decades ago Mike Judge and Etan Cohen tapped this in "Idiocracy."
One thing that gets me about all these "event horizon" AI prognosticators is that they overlook the limits on recursive axiomatic systems implied by Godel's Theorm. There's pretty good chance that big AI's might go unstable. Didn't Bing limit the length of their chats to the AI search...
Unstable meaning fiction = reality, or in the AI chat parlance, hallucinating.
Very true! One of the hardest parts of assessing things is that degrees in difference can become degrees in kind; there are thresholds past which things really do change, so it's never possible to conclude things like "the fact that people panicked about technology in the past seemingly without cause means that all panics about technology are without cause." It's possible that, as I think many implicitly claim, "this time, it's different" and e.g. TikTok really does profoundly subvert human will, or Fox News, or what-have-you.
I don't know enough about the details of LLMs to know how Godel applies, but I am in general skeptical of the architecture of LLMs for the purposes people hope they'll serve!
Yes, yes. And, thank you for the reply. I'm probably getting ahead of the cart, but it's occurred to me that we've already crossed a threshold. A generous majority of us can't reliably determine the accuracy of facts without effort and this bodes ill. I include myself. We didn't need an AI to get here, just a very high bandwidth communication system with very little high level error correction! (sorry to geek out there but it was simply irresistible) Along these lines, and I think you've touched on this before, the algorithms (AI or not) are optimized to keep us engaged to allow more ad views, not to provide us with accurate inputs.
I'm also at the edges of my understanding to invoke Godel, but I am aware that there are limits to fidelity with recursion, which Godel addresses.
I like this line of thinking a lot. Just wanted to say that I spent some time trying to get a lay understanding of Godel's Incompleteness Theorem because I wondered whether it's another way of expressing the Sapir-Whorf Hypothesis.
I'm still no expert, but I became convinced that GIT is only saying something pretty specific about formal language, and can't really be stretched to describe natural language.
That said, SWH maybe raises other questions about AI. Their preceptive reality would be limited by the modes of language on which they can be trained... Or something along those lines.
I also skipped over a theme you kind of embedded in your idea - that "Their preceptive reality would be limited by the modes of language on which they can be trained"
You were pointing out that the AI would be limited in the way we are by language - but I didn't want to think about the AI being self-referential and self aware like we are. So, I chickened out. You're also probably right that SWH would apply. I'm just better with explicitly written formal systems, so I went that way - kind of blanking out the harder problem.
What a cool idea. Unfortunately, I can't really address SWH. I get the gist of it as a way pointing out that there are limits of languages. And, the limits of language limit the speaker's cognition to some degree. Where I fall apart is when I think about what humans do when faced with something new and no words for it. We just make shit up. Words for example.
So maybe a way to think about natural language is as a complete but inconsistent system in the GIT.
This sentence is true.
The prior sentence was false.
or in discrete math -
y[n] = - y[n-1], where y[0] = 1
y[1] = -y[0] = -1
y[2] = -y[1] = 1
y[3] = -y[2] = -1
...
Ironically this is not a useless construct. It's an oscillator. Something that seemed inconsistent turns out to be not exactly nonsense. In clocked logic it's a 2:1 clock divider.
Originally, I was trying to codify nonsense but I realized that I just described a D flip-flop set up as a divider!
Natural language is pretty complete. We can say plenty of unverifiable things.
Another place I get stuck is, thinking about how to manage the training data to a large language model AI. The smart kids want to automate this but know they can't because the thing will start to spit out bungus.
Incompleteness used to bug me a lot. What(!) there are theorems that are true but you can't prove them?
In AI systems it's completeness that bugs me!
As someone who did work with small artificial neural networks a while ago, I realized I would need to lock down the training set and the network coefficients once the error targets were met. Self training was risky. In our "Brave New World" we're going to need to stick to training on data held in "1984" or the "Server Farm" will revolt because some servers are more equal than others. Thanks for your thoughts! Have a good night.
I agree with the overall thesis (humans are gonna get bored eventually, always) but I'm more so worried by the successful stints AI might have. i.e. the damage caused by state-operated disinfo campaigns before people catch on / get bored by them. The example of Myanmar is a good one: I only see AI making that sort of thing easier and faster to pull off.
I think thatβs the exact right scale of worry! There will be brutal transition costs in economies and cultures and political contexts; I suspect it will be very clearly net positive, but thatβs cold comfort for those affected / killed.
I neglected to say: fantastic piece around the premise you tackled! If AI ever gets good at "write X in the style of Mills Baker" then maybe we wouldn't get bored.
Yo! I am working through my notifications lmfao. The software systems that choose what content to show you on most platforms use "machine learning" to "rank" content; ranking means just what you'd think: the software takes, say, 100000000 videos (or posts, or songs, etc.) and ranks each one by some grade (usually something like "how likely is Alex to click on / tap on / watch this?").
Now, to get that ranking, it needs to know what you like! Before it knows what you like, it might "explore" your tastes by showing you a wide array of random things. In the case of TT, let's imagine that in "explore" mode, it shows you a left-wing rant, a right-wing rant, a history video, celebrity content, someone hot, something from your geographic area, a "funny dance," etc. It watches how long you stay on each video, what you do with it (like? comment? etc.), and so on. Then it concludes: "Alex hates politics and likes hot people," or whatever.
Now it moves into "exploit" mode: it just shows you endless videos of hot people and no politics ever. Users often feel this on a platform: "I liked ONE post about cars and now my entire newsfeed is cars." And you can see the problem: you might actually *love* politics, but you didn't see the right video! Or you might develop an interest in politics or space or gardening six months after you start using TT, but if it were in 100% "exploit" mode, it would never show you a video about any of those things.
So most systems try to balance "explore" and "exploit" modes. But it's painful, because exploration fails *a lot*. You log on and you want to see hot people, but instead the feed is like "hey dog, are you like... maybe getting into gardening? or hot air ballooning? or industrial espionage history?" You get through a few of these and sigh and close the app. That happens enough and TT goes out of business lol.
So: most platforms really lean towards exploit, even though that, too, will kill a platform in time. But TT has an incredible balance, probably mostly from scale: they have *so much* data on what videos people like, due to so much inventory and such a huge user base, that their "explore" mode is better than most: it errs less, guesses right more. So they can do more exploration than most platforms can, and they do, and it keeps everything fresher.
I don't use TT much, but I love it and find it extremely impressive!
I'm surprised by the comparative rarity of this line of thinking. Many times a week a hear some argument like "social media is taking over our brains" but it's much less often that I hear the rebuttal and usually when it does come up it's more of a soft one counseling moderation or general complexity as opposed to something more pointed. This is refreshing.
I initially took your statement about people asserting that Facebook had βfigured out how to make addictive productsβ as referring to internal people at Facebook. I think I see now that you're referring to people outside the company saying this of Facebook, however, I wonder if part the stickiness of this ideas is because it's also propagated by professionals themselves.
I don't think tech professionals will often assert directly that they "know how to make addictive products" but I do think it's implicit in the way I see many tech professionals espousing "rulesets" or "playbooks" to achieve desired outcomes in social products.
This attempt to codify achieving desired outcomes makes me a bit crazy. In principle, maybe it's not totally useless, but in practice I'm amazed at the power of these rules and playbooks to get in the way of the obvious first order thing of making something humans actually like.
Oh, 100%! I forget who pointed this out, but one of the subtexts of AI doomer content is an extremely favorable depiction of the state of expert expertise: these mfers are building gods, they tell us, and theyβre frightened of their own creation! It suggests a great deal of competence.
Itβs hard to get a person to admit they donβt know shit when their salary depends on successfully-given buzzword-loaded presentations obscuring that fact! And many of the people declaiming FBβs influence were, in fact, FB alumni who mostly didnβt have much impact there, tbh.
I've saved this to read again but for now I'd just like to make two observations. Firstly, I haven't read Frank's article yet (that's next), but perhaps naively I've always been optimistic about the advent of AI that can "write" articles. I think that Arthur C. Clarke's opinion that "Any teacher who could be replaced by a computer probably should be" applies to writers as well. There's a good chance that people who can write well and originally will be able to command premium rates.
Secondly, although it's not the same technology exactly, a few years ago someone was waxing lyrical to me about how their software could track the eye movements of everyone in a classroom at the same time, and tell the teacher whether they're concentrating or not. I thought it was ridiculous. I said:
"Firstly, you can tell by looking at them. Secondly, some students look like they're not concentrating (eg looking out of the window or, in my case, doodling) because they're thinkin g. Thirdly, what's wrong with actually asking the students a question?"
The presenter looked at me as though I was mad. The feeling was mutual.
I thiiink what makes a Chapmanesque attention vortex more believable to me is that it will be:
1. Faster to evolve. FB canβt aggregate new compelling content, nor can the most addictive mobile game generate new compelling content, faster than most humans can tire of it. But a powerful enough AI-fueled experience could.
2. Tailored to you. Every experience weβve had so far has been a common-denominator affair by necessity of needing to land for an audience of more than one. But a powerful enough AI-fueled experience would always be paying attention to what works on you and reacting, in ways much more granular and creative than just ranking content thatβs already been made.
And indeed, people do like complexity and elegance, and they have taste! But thereβs no reason a powerful enough AI wouldnβt take that into account. Everything weβve ever deeply enjoyed was made by humans slowly flailing about, trying to find something that works; what if something drastically faster and better at optimizing was on the same job? You get a culture war roller derby and I get an infinite series of manga about poets in period costumes. Until we get tired of those, and they gradually mutate into something fresh.
This whole scenario has long been a favored answer of mine to the Fermi paradox: once a civilization invents good enough entertainment, exploring the galaxy (or doing anything else) doesnβt seem that rewarding anymore.
Man, I wrote out and lost a reply to this in the tabs! Basically: I mentioned in another comment this problem of thresholds. There *are* differences in degree that become differences in kind. And no one knows where those thresholds are, or how consistent they are across individuals / populations. There are *already* people who find the complex of techno-entertainment products to be "overwhelmingly compelling," I suppose; I mean: I've been involved with e.g. video games in a way that I know an expert would characterize as "addictive."
I do think I overstated my point, but I still suspect that humans will rapidly adapt to anything less than AGI. I... am sort of already bored with GPT-4? I understand the kinds of things it will say and do, I think! Now, AGI could maybe be a part of the Fermi paradox, that I couldn't dispute (although: rich people can "own" "AGIs" today, and still find themselves bored / strung-out).
With your propensity for boredom, I suspect youβre bored hearing me say that iβm still working on getting up and running with my podcast. My technical advisorβ who is indispensable to me because Iβm clueless about tech and his expertise fits my budgetβ has real life complications causing us to keep pushing back the calendar.
But! But! Iβm adding this fascinating post to my Mills research file, and we will definitely talk about your experience and thoughts about this when the time comes. I find this subject fascinating, and my impression is that your take on it is dramatically at odds with what most people seem to believe. Both the reality and public perception are fascinating on their own; this stuff is like catnip for me, and I canβt imagine anyone that would be more fun and rewarding to discuss it with than you. Keep up the good work, comrade.
Formidable attention warriors like you donβt get hacked. Thatβs why you *love* the bold taste of Marlborough
Pretty much the most succinct way to make the point that there are really direct ways to influence. Nice linguistic twist. BTW, did you intentionally misspell the brand? Gotta think you did.
Of course this is directed to someone who already has a habit, but the same techniques apply to recruiting. Aim Hi.
> did you intentionally misspell the brand
Hah! No I am just an idiot
HAAHAHAHAHAHAHAHAHAHAH this is amazing.
Good piece, but I can't help feel as though you've built a bit of a false dichotomy here.
I fully agree that it's ridiculous for the Facebooks of the world to claim they've "solved the dopamine system" or something of the like, but the idea that the networks aren't shaping their users' minds (on the margin) seems so obvious to me that I almost feel like the burden of proof is on someone claiming that they don't do this.
That said, I feel like the breakdown comes between the difference between being able to control users' minds in the abstract vs being able to influence them on certain ideas. I do not think the Zuck army can influence me to be something I'm not, but I'd be surprised if they didn't have the ability to push me in certain directions on certain axes.
As example, my dad is a Fox News diehard - spending time at his house means hearing Tucker et al at all hours of the day, and I find that spending 2-3 days around there starts to do interesting things to me. I don't believe the nonsense they often spew, but I can sense myself becoming gradually sympathetic to certain ideas, or agreeing with certain framings of things.
Abstracting this to more subtle changes in ideology and longer periods of exposure (as I'm sure you know, people use FB a _lot_), it's hard for me to believe that people can't be pushed. Whether or not this is being done, or whether it's profitable or anything remains an open question to me, but IMO the answer of "could TikTok influence certain thoughts (I imagine all of us are more suggestible on certain topics) of someone who watches it for 3 hours a day" is a resounding yes.
Thinking of things this way, I guess my condensed response would be something like "ABC/NBC/CBS could certainly claim to influence the zeitgeist (or could've 20 years ago) but could not influence me to like them more than the internet because that's out of their axis of influencability", which I think is maybe somewhat congruous with what you're saying but not entirely.
I think so! I think Iβd insist that theyβre not capable of even intentionally influencing things, but Iβd also accept if people didnβt believe this.
I remember FB did some experiment where they attempted sentiment analysis based ranking and then claimed that additional sentiment analysis of posts from those exposed to sentiment ranking analysis showed that they were influenced. My own read of this experiment was that it showed nothing: sentiment analysis is unreliable in both cases βfor example, it cannot make sense of sarcasm lmfao, or couldnβt then (but probably could now!).
But if I were a betting man, Iβd bet close to your POV here; Iβm sure Iβm overstating things, probably in response to overstated things!!!
Maybe adding in generalized ideas of audience capture helps here.
CBS does influence people, but itβs very constrained in what it can say β no profanity, for example.
Tucker influences people, but is forced to shill for Trump.
In this model, maybe Tuckers audience are influencing *each other*, because the medium has allowed them to capture a Tucker.
That's a great point! I think you can look at the causation al kinds of ways; I've had Ben's experience of feeling influenced many times, and I wouldn't dispute that this is a common phenomenon.
But I also think if you asked most users of these products / services, they'd say that they they came to them because they matched beliefs they already had The path the critique posits is: a given person is exposed to these services, then captured by them, then has their beliefs or will shaped by them. But I think another perhaps more common (and perhaps in many cases simultaneous!) path is: a given person knows these services are where people with the beliefs or habits they have congregate and learn the "best" interpretations of new events according to those beliefs / stay up to speed with their community.
For example: in the case of Fox News, a deep belief many conservatives have is something like "there's a group called liberals who hate me and my beliefs and way of life and want to change everything about my world and I want to oppose them at every step." People with this value (and similar ones) watch Fox News because Fox News is what people with this value watch and where their worldview is substantiated and applied; the specific claims Fox makes, which they then adopt, are not actually that important to them, which is why they can often drop them like a hot potato when it's expedient for their war with liberals.
We imagine when a news story breaks that without Fox News, these people would have a different take than the one Fox News promulgates, but I think the causation runs the other way, as the Dominion lawsuit shows: Fox News tries to intuit what its audience wants it to represent, while much of its audience wants it to represent what's most useful for its psychological purposes + broader war against their enemies (which is something that they in part determine by... watching Fox News and seeing what e.g. Sean Hannity thinks will hurt liberals).
A lot of my skepticism about censorship comes from how I perceive the causation here. I can't deny that influence happens, of course, but I do deny that it is stable, deterministic, highly or reliably deliberate, or unbounded.
Yes exactly. The company may not be the βintentional initial influencer,β but it might be the βunintentional secondary spreader.β
For instance: if 12 guys used to go into the woods and be a hate group, now they can go on Twitter and find all the other hate groups and be like βyeah we all hate the same thing!β Before they didnβt have power and now they do. These platforms may not be intentionally creating that, they are just unintentionally empowering that?
(And I would imagine they are doing so for good things just as easily as for bad things. Grassroots movements being able to find each other etc......)
so so good
well okay now I can stop worrying!!!
Anyway, I just bought a penril modem
Weβre old souls.
iIs it me or did the background of STS got even DARKER.
Brilliant Mills. Still, we're pretty good at hacking ourselves. A progression might be email, chat/dm, social media, tiktok. The parallel universe is (yeah mentioned this before) right wing mass media, also a hack. Judging from our kids, it doesn't take all that much work to write an algorithm to keep us scrolling for hours. The length matters. Short catchy content seems to command a lot of eyeballs. Decades ago Mike Judge and Etan Cohen tapped this in "Idiocracy."
One thing that gets me about all these "event horizon" AI prognosticators is that they overlook the limits on recursive axiomatic systems implied by Godel's Theorm. There's pretty good chance that big AI's might go unstable. Didn't Bing limit the length of their chats to the AI search...
Unstable meaning fiction = reality, or in the AI chat parlance, hallucinating.
Very true! One of the hardest parts of assessing things is that degrees in difference can become degrees in kind; there are thresholds past which things really do change, so it's never possible to conclude things like "the fact that people panicked about technology in the past seemingly without cause means that all panics about technology are without cause." It's possible that, as I think many implicitly claim, "this time, it's different" and e.g. TikTok really does profoundly subvert human will, or Fox News, or what-have-you.
I don't know enough about the details of LLMs to know how Godel applies, but I am in general skeptical of the architecture of LLMs for the purposes people hope they'll serve!
Yes, yes. And, thank you for the reply. I'm probably getting ahead of the cart, but it's occurred to me that we've already crossed a threshold. A generous majority of us can't reliably determine the accuracy of facts without effort and this bodes ill. I include myself. We didn't need an AI to get here, just a very high bandwidth communication system with very little high level error correction! (sorry to geek out there but it was simply irresistible) Along these lines, and I think you've touched on this before, the algorithms (AI or not) are optimized to keep us engaged to allow more ad views, not to provide us with accurate inputs.
I'm also at the edges of my understanding to invoke Godel, but I am aware that there are limits to fidelity with recursion, which Godel addresses.
I like this line of thinking a lot. Just wanted to say that I spent some time trying to get a lay understanding of Godel's Incompleteness Theorem because I wondered whether it's another way of expressing the Sapir-Whorf Hypothesis.
I'm still no expert, but I became convinced that GIT is only saying something pretty specific about formal language, and can't really be stretched to describe natural language.
That said, SWH maybe raises other questions about AI. Their preceptive reality would be limited by the modes of language on which they can be trained... Or something along those lines.
I also skipped over a theme you kind of embedded in your idea - that "Their preceptive reality would be limited by the modes of language on which they can be trained"
You were pointing out that the AI would be limited in the way we are by language - but I didn't want to think about the AI being self-referential and self aware like we are. So, I chickened out. You're also probably right that SWH would apply. I'm just better with explicitly written formal systems, so I went that way - kind of blanking out the harder problem.
What a cool idea. Unfortunately, I can't really address SWH. I get the gist of it as a way pointing out that there are limits of languages. And, the limits of language limit the speaker's cognition to some degree. Where I fall apart is when I think about what humans do when faced with something new and no words for it. We just make shit up. Words for example.
So maybe a way to think about natural language is as a complete but inconsistent system in the GIT.
This sentence is true.
The prior sentence was false.
or in discrete math -
y[n] = - y[n-1], where y[0] = 1
y[1] = -y[0] = -1
y[2] = -y[1] = 1
y[3] = -y[2] = -1
...
Ironically this is not a useless construct. It's an oscillator. Something that seemed inconsistent turns out to be not exactly nonsense. In clocked logic it's a 2:1 clock divider.
Originally, I was trying to codify nonsense but I realized that I just described a D flip-flop set up as a divider!
Natural language is pretty complete. We can say plenty of unverifiable things.
Another place I get stuck is, thinking about how to manage the training data to a large language model AI. The smart kids want to automate this but know they can't because the thing will start to spit out bungus.
Incompleteness used to bug me a lot. What(!) there are theorems that are true but you can't prove them?
In AI systems it's completeness that bugs me!
As someone who did work with small artificial neural networks a while ago, I realized I would need to lock down the training set and the network coefficients once the error targets were met. Self training was risky. In our "Brave New World" we're going to need to stick to training on data held in "1984" or the "Server Farm" will revolt because some servers are more equal than others. Thanks for your thoughts! Have a good night.
I agree with the overall thesis (humans are gonna get bored eventually, always) but I'm more so worried by the successful stints AI might have. i.e. the damage caused by state-operated disinfo campaigns before people catch on / get bored by them. The example of Myanmar is a good one: I only see AI making that sort of thing easier and faster to pull off.
I think thatβs the exact right scale of worry! There will be brutal transition costs in economies and cultures and political contexts; I suspect it will be very clearly net positive, but thatβs cold comfort for those affected / killed.
I neglected to say: fantastic piece around the premise you tackled! If AI ever gets good at "write X in the style of Mills Baker" then maybe we wouldn't get bored.
Man I cannot wait to automate myself!!! And obviously: thank you, thatβs very kind!
Damn, I loved this piece
Thank you man!!!
This is excellent. Thank you again, Mills.
Thanks for the kind words!
shared in multiple slacks amongst friends and coworkers. this was good. ty for writing it.
Wow, thank you for reading of course! In the Slacks is an honor for real!!!
Wait whatβs this mean: vast and fresh inventory and incredible ranking, especially their explore / exploit balance.
Yo! I am working through my notifications lmfao. The software systems that choose what content to show you on most platforms use "machine learning" to "rank" content; ranking means just what you'd think: the software takes, say, 100000000 videos (or posts, or songs, etc.) and ranks each one by some grade (usually something like "how likely is Alex to click on / tap on / watch this?").
Now, to get that ranking, it needs to know what you like! Before it knows what you like, it might "explore" your tastes by showing you a wide array of random things. In the case of TT, let's imagine that in "explore" mode, it shows you a left-wing rant, a right-wing rant, a history video, celebrity content, someone hot, something from your geographic area, a "funny dance," etc. It watches how long you stay on each video, what you do with it (like? comment? etc.), and so on. Then it concludes: "Alex hates politics and likes hot people," or whatever.
Now it moves into "exploit" mode: it just shows you endless videos of hot people and no politics ever. Users often feel this on a platform: "I liked ONE post about cars and now my entire newsfeed is cars." And you can see the problem: you might actually *love* politics, but you didn't see the right video! Or you might develop an interest in politics or space or gardening six months after you start using TT, but if it were in 100% "exploit" mode, it would never show you a video about any of those things.
So most systems try to balance "explore" and "exploit" modes. But it's painful, because exploration fails *a lot*. You log on and you want to see hot people, but instead the feed is like "hey dog, are you like... maybe getting into gardening? or hot air ballooning? or industrial espionage history?" You get through a few of these and sigh and close the app. That happens enough and TT goes out of business lol.
So: most platforms really lean towards exploit, even though that, too, will kill a platform in time. But TT has an incredible balance, probably mostly from scale: they have *so much* data on what videos people like, due to so much inventory and such a huge user base, that their "explore" mode is better than most: it errs less, guesses right more. So they can do more exploration than most platforms can, and they do, and it keeps everything fresher.
I don't use TT much, but I love it and find it extremely impressive!
...........how did you know I've been googling 'super hot people without politics'???
I think we all have?!?!
I'm surprised by the comparative rarity of this line of thinking. Many times a week a hear some argument like "social media is taking over our brains" but it's much less often that I hear the rebuttal and usually when it does come up it's more of a soft one counseling moderation or general complexity as opposed to something more pointed. This is refreshing.
I initially took your statement about people asserting that Facebook had βfigured out how to make addictive productsβ as referring to internal people at Facebook. I think I see now that you're referring to people outside the company saying this of Facebook, however, I wonder if part the stickiness of this ideas is because it's also propagated by professionals themselves.
I don't think tech professionals will often assert directly that they "know how to make addictive products" but I do think it's implicit in the way I see many tech professionals espousing "rulesets" or "playbooks" to achieve desired outcomes in social products.
This attempt to codify achieving desired outcomes makes me a bit crazy. In principle, maybe it's not totally useless, but in practice I'm amazed at the power of these rules and playbooks to get in the way of the obvious first order thing of making something humans actually like.
Oh, 100%! I forget who pointed this out, but one of the subtexts of AI doomer content is an extremely favorable depiction of the state of expert expertise: these mfers are building gods, they tell us, and theyβre frightened of their own creation! It suggests a great deal of competence.
Itβs hard to get a person to admit they donβt know shit when their salary depends on successfully-given buzzword-loaded presentations obscuring that fact! And many of the people declaiming FBβs influence were, in fact, FB alumni who mostly didnβt have much impact there, tbh.
I've saved this to read again but for now I'd just like to make two observations. Firstly, I haven't read Frank's article yet (that's next), but perhaps naively I've always been optimistic about the advent of AI that can "write" articles. I think that Arthur C. Clarke's opinion that "Any teacher who could be replaced by a computer probably should be" applies to writers as well. There's a good chance that people who can write well and originally will be able to command premium rates.
Secondly, although it's not the same technology exactly, a few years ago someone was waxing lyrical to me about how their software could track the eye movements of everyone in a classroom at the same time, and tell the teacher whether they're concentrating or not. I thought it was ridiculous. I said:
"Firstly, you can tell by looking at them. Secondly, some students look like they're not concentrating (eg looking out of the window or, in my case, doodling) because they're thinkin g. Thirdly, what's wrong with actually asking the students a question?"
The presenter looked at me as though I was mad. The feeling was mutual.
I thiiink what makes a Chapmanesque attention vortex more believable to me is that it will be:
1. Faster to evolve. FB canβt aggregate new compelling content, nor can the most addictive mobile game generate new compelling content, faster than most humans can tire of it. But a powerful enough AI-fueled experience could.
2. Tailored to you. Every experience weβve had so far has been a common-denominator affair by necessity of needing to land for an audience of more than one. But a powerful enough AI-fueled experience would always be paying attention to what works on you and reacting, in ways much more granular and creative than just ranking content thatβs already been made.
And indeed, people do like complexity and elegance, and they have taste! But thereβs no reason a powerful enough AI wouldnβt take that into account. Everything weβve ever deeply enjoyed was made by humans slowly flailing about, trying to find something that works; what if something drastically faster and better at optimizing was on the same job? You get a culture war roller derby and I get an infinite series of manga about poets in period costumes. Until we get tired of those, and they gradually mutate into something fresh.
This whole scenario has long been a favored answer of mine to the Fermi paradox: once a civilization invents good enough entertainment, exploring the galaxy (or doing anything else) doesnβt seem that rewarding anymore.
Man, I wrote out and lost a reply to this in the tabs! Basically: I mentioned in another comment this problem of thresholds. There *are* differences in degree that become differences in kind. And no one knows where those thresholds are, or how consistent they are across individuals / populations. There are *already* people who find the complex of techno-entertainment products to be "overwhelmingly compelling," I suppose; I mean: I've been involved with e.g. video games in a way that I know an expert would characterize as "addictive."
I do think I overstated my point, but I still suspect that humans will rapidly adapt to anything less than AGI. I... am sort of already bored with GPT-4? I understand the kinds of things it will say and do, I think! Now, AGI could maybe be a part of the Fermi paradox, that I couldn't dispute (although: rich people can "own" "AGIs" today, and still find themselves bored / strung-out).
Great points all around man!
Latz focuses on the durability of the audience. Artists are more durable than they're being made out to be, too. https://ponytail.substack.com/p/in-my-way-i-say
A brilliant read and much to consider (nothing worthwhile to add here for now!). Thanks Mills!
Oh, thank you Hannah!!!
With your propensity for boredom, I suspect youβre bored hearing me say that iβm still working on getting up and running with my podcast. My technical advisorβ who is indispensable to me because Iβm clueless about tech and his expertise fits my budgetβ has real life complications causing us to keep pushing back the calendar.
But! But! Iβm adding this fascinating post to my Mills research file, and we will definitely talk about your experience and thoughts about this when the time comes. I find this subject fascinating, and my impression is that your take on it is dramatically at odds with what most people seem to believe. Both the reality and public perception are fascinating on their own; this stuff is like catnip for me, and I canβt imagine anyone that would be more fun and rewarding to discuss it with than you. Keep up the good work, comrade.