It would be hard to build a cross platform humane recommendation algorithm for specific content. But I think it would be possible to build a global algorithm that ranks trends themselves. A sort of ranked google trends so at least I know what to search for.
Why would it be so hard? It SEEMS like it should be possible to, say, take all the text I've written and predict with quite some accuracy if I'd be interested in a given document. (Especially if I provide some additional supervision.) The fact that no such algorithm already exists seems like pretty decent evidence that it is in fact hard, but I'm just wondering why...
That's super interesting! I've never heard of recommendation / ranking algorithms using "things I've written" as a signal. Seems like that could be a *very* rich source. Of course it has limitations, there are lots of things I'm interested in reading about for fun that I ~never write a word about, but still seems huge.
(Disclaimer: I don't know much about recommendation / ranking algorithms.)
Yeah, I'd think it would be a very strong source! My guess is an algorithm could probably predict all the places I've lived, simply on the basis of little quirks in word choice. So while I'd love to believe that I have interests that AI wouldn't know about, my guess is that an ideal AI would do unnervingly well in predicting those too. And in any case, a huge fraction of a good recommendation is probably tone/style/attitude as much as topic? And that's all probably VERY legible?
I think it’d be hard for a single algorithm to index all of YouTube, Instagram, Reddit, and give you a single feed with full depth. Maybe you could have an agent in the background trying various searches based on your preferences and building you a custom feed?
I totally agree that one’s own writing (and reading) is a great implicit filter. Not only for discovering content, but also to style it according to our preferences.
> I think it’d be hard for a single algorithm to index all of YouTube, Instagram, Reddit, and give you a single feed with full depth. Maybe you could have an agent in the background trying various searches based on your preferences and building you a custom feed?
Yup. This is "AI assistant" territory - a slam dunk when we all have a Phd-smart, maximally conscientious AI assistant who can learn our individual tastes in exquisite detail, and THEN can curate media feeds for us, breaking free entirely of external algorithms.
"but if all the information that enters your brain is being filtered by an algorithm, it seems important that you know the algorithm is on your side." Seems far too late for that to me and even unlikely that this could ever have been an option. But I am trying to claw away from a pit of despair this week, feeling like Sam when he realizes that Frodo can't destroy the ring.
Don't despair! There are reasons to think thinks could get better! For one thing, we know have a much better idea what algorithmic ranking leads to—culture can now prepare the appropriate adaptations. Also, and this is speculative, but maybe one of the reasons current algorithms are so enshittified is that there are such strong network effects, so established companies can really turn the screws once they've gotten hold. But if AI can do all the things the crowd used to do, maybe you've got more of a one-sided marketplace and it's more competitive?
From a wise professor in the old timey days: A king once sentenced a man named John to die for some crime and after the sentence was proclaimed, John said, if I can teach your horse to sing in 6 months, will you pardon me? The king, smirking and snearing, agreed. A friend, overhearing this said to John, you are crazy! You can't teach a horse to sing! John said: in six months, the king might die. I might die. Or, the horse may learn to sing.
I'm finding that I'm the kind of person who gets 'done' with topics. Perhaps I've learned what I wanted, or accomplished my task at hand, or often It just seems that I've "bottomed out" and reached the point of diminishing returns on the material the algorithm is suggesting for me. I need to move on. I'd like to see an anti-algorithm as part of every suggestion matrix - a few tildas, naught symbols, and brackets in the coding that would suggest material that is the diametric opposite of what I have been focusing on. Yeah, it would be hard to convince the source media - I can't think of any incentives doing it this way would provide. But maybe we could have third party applications that would try to do this for you. Say, click a button and get antiYouTube. A dream, I know, and entirely contrary to human nature.
I want to be on the internet to see cats. I don't want to be advertised to or told to watch something else. Now I can't do that. I'm forced to either give in or stop.
I think I might get a coffe at the local cat cafe.
I can't do group chats. That drives me crazy. I tried discord. It was too much. Maybe cause I'm old. But I'm not the target audience. I do feel for the kids today. Being bombarded is all they've ever known. It's hard not to get sucked in.
This got me thinking about the “long tail” of the Internet, which still exists, and what you seem to be asking for is Google’s search engine back when they zealously guarded against gaming the results.
Personally, I’m OK with finding new stuff from links of other people I read — it’s how I ended up here! But taking a step back, I grew up in a very different environment, pre-Internet (b 1971). Sometimes I’ll be in the throes of scrolling, but more likely I’ll read something, ask myself a question, and search. I wonder if that style of curiosity can continue in any algorithmically dominant present and future.
Not trying to shill, but Kagi (dot com) up-rank search results from "small web" (non-commercial / personal) sites, and down-rank results from sites with a bunch of advertising and tracking. It feels _kind of_ like using Google from 15 years ago.
I didn't! Being 100% psychologically incapable of watching videos prevents me from testing it in detail, but I like where it's coming from. (Suspect automation will greatly help efforts like this in the future.)
I—well—I more or less can't watch videos for the purposes of getting information. I mean, I *can*, but I hate it and I almost never do it. I infinitely prefer text because you can so easily zoom around in the content. To me, text feels more like an "open world" and video feels like being on "rails".
(I like movies and audiobooks just fine. I guess I don't prefer strongly prefer novels because there's not much reason to skim around in a novel?)
The solution offered in the second one in particular suggests an approach which is not wholly disconnected from your idea, with an LLM in there to do the curation:
“LM agents would obviously not have to be narrowly behaviourist. Instead, they would understand our preferences and (societal) values in the same way that we do: they would be able to analyse and deploy the concepts that we attach to them. They already display impressive facility at this task [58]; this can be expected to improve as they improve in other dimensions. An LM recommender could talk with the user about what they want to see, both in the abstract and through dialogical engagement about specific recommendations [73]. Training it would be similar to a process of training a human assistant, giving direct instructions from which they immediately learn, so that they can instinctively operate without supervision. This would be a huge benefit to us! Instead of simply propagating our revealed preferences (and often pandering to our worst selves), recommenders could engage in a thoughtful, quasi- empathetic dialogue about how we want to allocate our attention. They can also nudge us towards better choices, and if necessary also incorporate societal values to guard against our individually rational choices having collectively irrational consequences [12].”
Not sure that sounds like a marginal-user use case, now that I read it back to myself. But maybe it's time to ping Seth Lazar and join forces with a shadowy cartel of user-centric free-range content recommendation?
I would have never thought that these issues could have been considered part of ethics! But when you read the papers it seems to make sense. The second paper is indeed VERY close to what I was suggesting, which I would have never suspected from the title. Maybe this is an idea that's somehow "in the water" now?
It's also interesting to see the same kinds of ideas reframed in such unusual ways, e.g. I pondered this sentence: "Others have already argued that digital technologies in general can have a deskilling effect on attention allocation."
I've been waiting for someone to clearly and forcefully make this point. Well done!
The software we want and need and deserve cannot be reached using the dominant status quo business models. But it could be reached with innovative business models that figure out how to rigorously pre-commit to certain constraints. And, if reached, it could make some people very rich, and everyone better off. Let's do this!
1. The sliders exist! At risk of shilling 2x in the same comment section, Kagi lets you promote, demote, and block specific sites in your search results, and it durably remembers your preferences. They have a public search stats page with a "Domain insights" section showing the most-blocked domains across their user base. (Pinterest dominates the list.) This is a user-paid subscription service (they don't harvest eyeball-milliseconds for advertisers), but they have the same basic incentive (attract and retain subscribers) as Netflix.
2. User-supplied content curation algorithms are part of the vision presented in Mike Masnick's "Protocols, Not Platforms" essay, and an early objective of various fediverses (e.g. Mastodon and Bluesky). But even on content platforms with an open protocol or API, the folks in my 99th-percentile-tech-aware circles don't tend to tweak their own sliders in the way you describe. People are putting RSS feeds into tools like Minifeed, and following / blocking accounts on Bluesky, not much different than 20 years ago. 🤷
3. Is it feasible to make self-actualization more legible to capitalism? (Not a rhetorical question! I don't know the answer.)
If someone can serve content that helps me live according to my most deeply-held values and reach my long-term goals, I absolutely want to reward them. But, I don't feel like a consistently good reward allocator, as I'm prone to losing sight of my own values and goals. For example, I _feel_ like Dynomight internet content helps me understand the world better, and I trust the intentions of the author. But maybe I'm wasting time on someone's blog when I could be doing a professional development task or lifting weights.
So, I fear that the answer is "probably not". But then I remember things like Income Share Agreements, where a school is directly exposed to students income after graduation. I also imagine schemes such as: paying someone to follow me around and ask "WTF are you doing, man?" every time I pull out my phone, and they get a bonus if I self-rate my day as well-spent.
4. To an extent, LLMs can serve this need effectively, especially ones that will search the web for you. "Recommend some rhythm-oriented piano jazz to play while I work. Something at least an hour long with no vocals."
> What exactly should those structures be? And what exactly is “control”, after all? I don’t know! Those seem like difficult technical problems. But they don’t seem that hard, do they? I suspect the main reason they haven’t been solved is that we haven’t tried very hard. We should do that.
The main problem here is that if you're not letting "economics" do it (which has created the current algorithms), the only other lever is "regulation," where if you're American, you're going to have a collection of 80+ year old half-corpse multi-millionaires who neither use nor understand the internet deciding on the answers to all of these questions, while heavily "advised" by lobbyists from all the FAANGS.
Do you REALLY think that's going to work out better?
But those are basically the only two options.
The real solution, of course, is going to be personal AI assistants that curate our media consumption streams. Their personalized algorithm will replace all the others.
Economics is the option! But economics only creates things if people want them. I think the idea of "demand non-revocable control of the utility function" isn't an entirely obvious idea, and the benefits wouldn't have been obvious 20 years ago, but now is the time.
Agree with your last paragraph of course, but you do still have the issue of the utility function of the assistant. What makes me optimistic is it seems like there won't be the same issue of a 2-sided market where a giant company can seize control and then change the algorithm after.
I just wanted to say thank you for your blog! I really enjoy your topics and writing style.
I guess I’m lucky because my behavior happens to create incentives for rankings that align well with my goals in life. Still, I have a thought:
Maybe people should learn algorithm literacy in school. Simply put, it’s not your goals but your actual actions—your active decisions put into practice—that shape your life.
It would be hard to build a cross platform humane recommendation algorithm for specific content. But I think it would be possible to build a global algorithm that ranks trends themselves. A sort of ranked google trends so at least I know what to search for.
Why would it be so hard? It SEEMS like it should be possible to, say, take all the text I've written and predict with quite some accuracy if I'd be interested in a given document. (Especially if I provide some additional supervision.) The fact that no such algorithm already exists seems like pretty decent evidence that it is in fact hard, but I'm just wondering why...
That's super interesting! I've never heard of recommendation / ranking algorithms using "things I've written" as a signal. Seems like that could be a *very* rich source. Of course it has limitations, there are lots of things I'm interested in reading about for fun that I ~never write a word about, but still seems huge.
(Disclaimer: I don't know much about recommendation / ranking algorithms.)
Yeah, I'd think it would be a very strong source! My guess is an algorithm could probably predict all the places I've lived, simply on the basis of little quirks in word choice. So while I'd love to believe that I have interests that AI wouldn't know about, my guess is that an ideal AI would do unnervingly well in predicting those too. And in any case, a huge fraction of a good recommendation is probably tone/style/attitude as much as topic? And that's all probably VERY legible?
I think it’d be hard for a single algorithm to index all of YouTube, Instagram, Reddit, and give you a single feed with full depth. Maybe you could have an agent in the background trying various searches based on your preferences and building you a custom feed?
I totally agree that one’s own writing (and reading) is a great implicit filter. Not only for discovering content, but also to style it according to our preferences.
> I think it’d be hard for a single algorithm to index all of YouTube, Instagram, Reddit, and give you a single feed with full depth. Maybe you could have an agent in the background trying various searches based on your preferences and building you a custom feed?
Yup. This is "AI assistant" territory - a slam dunk when we all have a Phd-smart, maximally conscientious AI assistant who can learn our individual tastes in exquisite detail, and THEN can curate media feeds for us, breaking free entirely of external algorithms.
But tbh, that's just around the corner.
"but if all the information that enters your brain is being filtered by an algorithm, it seems important that you know the algorithm is on your side." Seems far too late for that to me and even unlikely that this could ever have been an option. But I am trying to claw away from a pit of despair this week, feeling like Sam when he realizes that Frodo can't destroy the ring.
Don't despair! There are reasons to think thinks could get better! For one thing, we know have a much better idea what algorithmic ranking leads to—culture can now prepare the appropriate adaptations. Also, and this is speculative, but maybe one of the reasons current algorithms are so enshittified is that there are such strong network effects, so established companies can really turn the screws once they've gotten hold. But if AI can do all the things the crowd used to do, maybe you've got more of a one-sided marketplace and it's more competitive?
From a wise professor in the old timey days: A king once sentenced a man named John to die for some crime and after the sentence was proclaimed, John said, if I can teach your horse to sing in 6 months, will you pardon me? The king, smirking and snearing, agreed. A friend, overhearing this said to John, you are crazy! You can't teach a horse to sing! John said: in six months, the king might die. I might die. Or, the horse may learn to sing.
There's some {John, king, horse} ↔ {dynomight, AI, algorithmic ranking} correspondence that's eluding me here. (Probably that means I'm the horse.)
I'm finding that I'm the kind of person who gets 'done' with topics. Perhaps I've learned what I wanted, or accomplished my task at hand, or often It just seems that I've "bottomed out" and reached the point of diminishing returns on the material the algorithm is suggesting for me. I need to move on. I'd like to see an anti-algorithm as part of every suggestion matrix - a few tildas, naught symbols, and brackets in the coding that would suggest material that is the diametric opposite of what I have been focusing on. Yeah, it would be hard to convince the source media - I can't think of any incentives doing it this way would provide. But maybe we could have third party applications that would try to do this for you. Say, click a button and get antiYouTube. A dream, I know, and entirely contrary to human nature.
> diametric opposite of what I have been focusing on
"You've been reading a lot of RCTs recently. Why not study some medieval witchcraft?"
I want to be on the internet to see cats. I don't want to be advertised to or told to watch something else. Now I can't do that. I'm forced to either give in or stop.
I think I might get a coffe at the local cat cafe.
I suspect a lot of the traditional cat action has migrated to group chats? (Arguably, group chats are the true non-algorithmic social media today?)
I can't do group chats. That drives me crazy. I tried discord. It was too much. Maybe cause I'm old. But I'm not the target audience. I do feel for the kids today. Being bombarded is all they've ever known. It's hard not to get sucked in.
This got me thinking about the “long tail” of the Internet, which still exists, and what you seem to be asking for is Google’s search engine back when they zealously guarded against gaming the results.
Personally, I’m OK with finding new stuff from links of other people I read — it’s how I ended up here! But taking a step back, I grew up in a very different environment, pre-Internet (b 1971). Sometimes I’ll be in the throes of scrolling, but more likely I’ll read something, ask myself a question, and search. I wonder if that style of curiosity can continue in any algorithmically dominant present and future.
Not trying to shill, but Kagi (dot com) up-rank search results from "small web" (non-commercial / personal) sites, and down-rank results from sites with a bunch of advertising and tracking. It feels _kind of_ like using Google from 15 years ago.
OK, just signed up for Kagi and suddenly find that *I* am the marginal user!
Update: Searching for "are ultrasonic humidifiers safe" gives dynomight.net/humidifiers as the second result. Sold!
Do you know the Tournesol project: https://tournesol.app/
I didn't! Being 100% psychologically incapable of watching videos prevents me from testing it in detail, but I like where it's coming from. (Suspect automation will greatly help efforts like this in the future.)
What does that mean?? :o I'm very curious, in case you're willing to share
I—well—I more or less can't watch videos for the purposes of getting information. I mean, I *can*, but I hate it and I almost never do it. I infinitely prefer text because you can so easily zoom around in the content. To me, text feels more like an "open world" and video feels like being on "rails".
(I like movies and audiobooks just fine. I guess I don't prefer strongly prefer novels because there's not much reason to skim around in a novel?)
This sounds just like StumbleUpon, RIP.
I remember you could choose subjects and rate each page you landed on, which would influence the next pages you’d see.
Never used it. Seems like a strange story (https://en.wikipedia.org/wiki/StumbleUpon). I guess ebay just kinda strangled it into irrelevance?
You know what this friendly and approachable post needs? Links to some dry, dense academic papers on morality!
These two are both about recommendations as an ethical problem:
* https://philpapers.org/rec/SCHAMS-10
* http://arxiv.org/abs/2410.12123
The solution offered in the second one in particular suggests an approach which is not wholly disconnected from your idea, with an LLM in there to do the curation:
“LM agents would obviously not have to be narrowly behaviourist. Instead, they would understand our preferences and (societal) values in the same way that we do: they would be able to analyse and deploy the concepts that we attach to them. They already display impressive facility at this task [58]; this can be expected to improve as they improve in other dimensions. An LM recommender could talk with the user about what they want to see, both in the abstract and through dialogical engagement about specific recommendations [73]. Training it would be similar to a process of training a human assistant, giving direct instructions from which they immediately learn, so that they can instinctively operate without supervision. This would be a huge benefit to us! Instead of simply propagating our revealed preferences (and often pandering to our worst selves), recommenders could engage in a thoughtful, quasi- empathetic dialogue about how we want to allocate our attention. They can also nudge us towards better choices, and if necessary also incorporate societal values to guard against our individually rational choices having collectively irrational consequences [12].”
Not sure that sounds like a marginal-user use case, now that I read it back to myself. But maybe it's time to ping Seth Lazar and join forces with a shadowy cartel of user-centric free-range content recommendation?
P.S. I scrapbooked this: https://danmackinlay.name/notebook/recommender_dynamics.html
I would have never thought that these issues could have been considered part of ethics! But when you read the papers it seems to make sense. The second paper is indeed VERY close to what I was suggesting, which I would have never suspected from the title. Maybe this is an idea that's somehow "in the water" now?
It's also interesting to see the same kinds of ideas reframed in such unusual ways, e.g. I pondered this sentence: "Others have already argued that digital technologies in general can have a deskilling effect on attention allocation."
Together with some rather smart people, a few years ago I wrote a paper about AI loyalty that touched on these issues.
https://www.peterbartreiner.com/uploads/1/2/0/0/120094820/aguirre-2020.pdf
I've been waiting for someone to clearly and forcefully make this point. Well done!
The software we want and need and deserve cannot be reached using the dominant status quo business models. But it could be reached with innovative business models that figure out how to rigorously pre-commit to certain constraints. And, if reached, it could make some people very rich, and everyone better off. Let's do this!
1. The sliders exist! At risk of shilling 2x in the same comment section, Kagi lets you promote, demote, and block specific sites in your search results, and it durably remembers your preferences. They have a public search stats page with a "Domain insights" section showing the most-blocked domains across their user base. (Pinterest dominates the list.) This is a user-paid subscription service (they don't harvest eyeball-milliseconds for advertisers), but they have the same basic incentive (attract and retain subscribers) as Netflix.
2. User-supplied content curation algorithms are part of the vision presented in Mike Masnick's "Protocols, Not Platforms" essay, and an early objective of various fediverses (e.g. Mastodon and Bluesky). But even on content platforms with an open protocol or API, the folks in my 99th-percentile-tech-aware circles don't tend to tweak their own sliders in the way you describe. People are putting RSS feeds into tools like Minifeed, and following / blocking accounts on Bluesky, not much different than 20 years ago. 🤷
3. Is it feasible to make self-actualization more legible to capitalism? (Not a rhetorical question! I don't know the answer.)
If someone can serve content that helps me live according to my most deeply-held values and reach my long-term goals, I absolutely want to reward them. But, I don't feel like a consistently good reward allocator, as I'm prone to losing sight of my own values and goals. For example, I _feel_ like Dynomight internet content helps me understand the world better, and I trust the intentions of the author. But maybe I'm wasting time on someone's blog when I could be doing a professional development task or lifting weights.
So, I fear that the answer is "probably not". But then I remember things like Income Share Agreements, where a school is directly exposed to students income after graduation. I also imagine schemes such as: paying someone to follow me around and ask "WTF are you doing, man?" every time I pull out my phone, and they get a bonus if I self-rate my day as well-spent.
4. To an extent, LLMs can serve this need effectively, especially ones that will search the web for you. "Recommend some rhythm-oriented piano jazz to play while I work. Something at least an hour long with no vocals."
> What exactly should those structures be? And what exactly is “control”, after all? I don’t know! Those seem like difficult technical problems. But they don’t seem that hard, do they? I suspect the main reason they haven’t been solved is that we haven’t tried very hard. We should do that.
The main problem here is that if you're not letting "economics" do it (which has created the current algorithms), the only other lever is "regulation," where if you're American, you're going to have a collection of 80+ year old half-corpse multi-millionaires who neither use nor understand the internet deciding on the answers to all of these questions, while heavily "advised" by lobbyists from all the FAANGS.
Do you REALLY think that's going to work out better?
But those are basically the only two options.
The real solution, of course, is going to be personal AI assistants that curate our media consumption streams. Their personalized algorithm will replace all the others.
Economics is the option! But economics only creates things if people want them. I think the idea of "demand non-revocable control of the utility function" isn't an entirely obvious idea, and the benefits wouldn't have been obvious 20 years ago, but now is the time.
Agree with your last paragraph of course, but you do still have the issue of the utility function of the assistant. What makes me optimistic is it seems like there won't be the same issue of a 2-sided market where a giant company can seize control and then change the algorithm after.
How about some form of a smart contract/blockchain-based system that wasn't alterable, and was designed with what you have in mind at its core?
Maybe! Although AFAIK we don't have a lot of examples of complex applications/algorithms running on the blockchain yet?
No, not yet. But it's just a matter of time.
I just wanted to say thank you for your blog! I really enjoy your topics and writing style.
I guess I’m lucky because my behavior happens to create incentives for rankings that align well with my goals in life. Still, I have a thought:
Maybe people should learn algorithm literacy in school. Simply put, it’s not your goals but your actual actions—your active decisions put into practice—that shape your life.