The following is a rough transcript which has not been revised by The Jim Rutt Show or Monica Anderson. Please check with us before using any quotations from this transcript. Thank you.
Jim: Today’s guest is Monica Anderson. She is CTO at Syntience Inc. And like very few other people around, she has over 20 years of hand-on experience with deep neural nets and related technologies. And I love this, she styles herself an experimental epistemologist.
What a great concept. As regular listeners to the show know, I’ll often say when I hear the word metaphysics, I reach for my pistol and all I need is a real world and some epistemology. So I like the combination of experimental epistemologist. And as she says in her own bio online, “There aren’t too many of those running around.” You can see more of what Monica’s up to at her website@experimentalepistemology.ai.
That’s quite a mouthful and quite a bunch to type down. But as usual, the link will be on our episodePage@jimrutshow.com. So check it out. Welcome Monica.
Monica: Hello. Welcome. Thank you. Thank you for having me. This sounds like a challenge and should be fun.
Jim: Yeah, I think it should be indeed. Today we’re going to mostly talk about her recent paper she published back in what October 2022 called Bubble City, Design Proposal for a Twitter Alternative, which is not a social media, it is a real-time idea router. And that kind of explains it at the highest level. And as I talked to, I told Monica in the pre-show chat, I’m seeing a lot of this around. This idea that we now finally have the tools into technology to in some sense let content route itself.
And this falls very strongly into that category of system ideas. So maybe Monica, if you’ll come back here, we can get started.
Monica: Sure. So I have had many decades of social media experience and message systems as such. I worked for chat company, wrote a chat client for the Macintosh among other things. And I like realtime information systems because there’s a lot more you can do in the realtime system than something that takes longer, like email. And how could we extend the idea of information management from being just something like web search to being a realtime thing?
And the way I envisioned doing that is to create a system which I call Bubble City. And the system is built on the principle that everything is done by pull. You have to request all the information that you want to use, that you want to hear. So you can say that one of the principles of Bubble City is that if you will claim the right to say whatever you want, then I will claim my right not to have to hear it.
And so we are creating a system where you are, with very simple means, very little effort can create something we can call an information bubble. And technically, it’s implemented as a filter for messages and it’s a very solid filter and it can sort, filter the message by topic because it is supported by an artificial intelligence of some kind, we will say.
There’s multiple levels of this. Back in 2007, I was working for Google and I had this an idea for something like Bubble City as something I wanted to do as a 20% project at Google. And I did that and afterwards Google wanted me to patent some of the things that I had come up with and so I did. And so parts of Bubble City were patented by Google in 2007.
I don’t think that’s going to be a problem, mainly because I use the Google search engine in reverse as the core of the filtering. And today we have at least two other solutions. We can use systems like ChatGPT’s APIs at the GPT level to do the filtering for us, or some other future AI that’s going to be cost effective.
And I have my own entry into this battle.
Jim: I got one, I want to kind of get this more tangible so people can get a feel for it. You know, describe it as a Twitter alternative. And so I think a lot of our listeners know what Twitter is. There’s vast sums, hundreds of millions a day of tweets being entered into the system. And we follow them by following people typically, right?
As I understand it in Bubble City, instead of following people, we would create or subscribe to existing bubbles that are defined topically, and they could be defined in as fine-grained a fashion as one wants. And essentially messages that fit the filter or the membrane pass through rules for the bubble enter into the bubble and you then see them. And you can then as the user of the bubble, add additional filters for what you want to see and how perhaps maybe you could order them. Is that approximately correct?
Monica: This is exactly correct. Right on. And maybe if I wanted to, I should put a various concrete use case, use as an example of where this could be useful. Consider that you are maybe are a living in New York and you want to go to a conference in Boston and you’re looking for [inaudible 00:05:33] or looking for opportunities to carpool to the conference. So you type in the name of the conference and the carpool and suddenly you’re in the chat room where people are talking carpooling to the conference. And some of the messages may be a few days old, some of them may be recent. And if you type something, a number of people who are in the same bubble because they entered the same search if you will, can immediately see your messages.
So we have created a virtual ephemeral chat room that is based on the topics of messages, which is handled and maintained by the artificial intelligence.
And this AI is very good at classifying stuff. So if it’s notices somebody is trying to sell sunglasses to the people going to the conference, the AI will notice that this is an attempt to sell stuff and we’re not going to let that message in, even though they’re talking about the conference.
And so this is a way basically to quench the social media flood of the information that we have, which is often a very low grade. And we can as professionals, for instance, we can go in and research only our specialty. We can create a chatroom, for instance, of all researchers that are using a system that are interested in small cell lung cancer. And you can basically exchange information in real time.
So the difference is it’s not a social media. You don’t follow people. Because on social media, people are generally used as a proxy for content ,because we didn’t know how to analyze the content. Now, we have the ability to do so and so let’s have our computers our AIs fully understand the content and only send the messages to the people who want to hear them.
So in this system, in Bubble City, you can type whatever you want with the legal limits of separate conversation, and only the people who want to see those messages will see them. It has to be a two level structure, which is that you use the messages that in, AI filtered messages to decide where to go. But once you’re in one of those chat rooms, your protection against spam and advertising and other things like hate goes down, and you may have to do additional filtering to get rid of that.
But those are some of the aspects of the system that make it radically different from what you see today in social media.
Jim: Okay. Now, you several times have talked about real time and what sounds like synchronous usage. Now, if we look at something like Twitter, only a very few people who use Twitter in a synchronous fashion. Most of us use it asynchronously. We check in once or twice a day, do our due, et cetera. And the realities of life being what they are, synchronous communications is really hard to make work for large groups of people while asynchronous is an easy way to make things scale.
So when you say synchronous or realtime, are we sure that’s what you really mean? Well
Monica: It is realtime with a log. So everybody who joins the bubble will see messages going back potentially months that other people have written. So there is context going back. It’s a very rarely discussed topic. You might basically open a window to it and set an alert and nothing happens for months and then somebody speaks to you. So there is definitely both aspects of it. I mean the realtime is important and I’m making it as much realtime as I can. It turns out that AI is going to take some time to understand the messages, but there’s going to be a few seconds and that will get better over time. But the realtime aspect is important because and one way to think about this is that it is a professionals’ brainstorming machine where professionals with a hard problem to solve can go get together anywhere on the planet and converse with each other potentially in realtime and solve these problems.
And as AIs get stronger, we’re going to introduce those AIs into this conversation first by individual use saying that I want my personal AI to be part of this, or later by somebody saying we’re going to have a AI for this bubble or something like that.
Jim: Yeah, one thing I’ve talked about for years is it would be great to have a librarian bot available. So if there’s a research question, you could pose it to the bot and say, “Bot, go research this, see what else is known about this conjecture.”
Monica: And now that we have things like ChatGPT and the next generation systems of this kind, we can see how this is going to be implemented. You’ll have your librarian, it’s going to be a subset of many other things that this bot’s going to do.
Jim: Yeah. Indeed. Now, so in some sense, this sounds a little bit like a worldwide slack system where could a bubble be thought to be analogous to a Slack channel?
Monica: But it has significant advantages. And one of these is a concept of addressability. If you’re a user in a system and you have a handle such as a Twitter handle or a name Slack, people can post stuff to you and you would’ve to explicitly ban them or quench them or whatever, silence them to avoid that.
But if you are a user and nobody can send messages to you, you don’t have this problem.
Jim: Yeah, good distinction. So bubbles are only content routing. So bubbles attract content that meet the profile of that bubble. For instance-
Monica: Correct.
Jim: … let’s say you’re a sports fan and are interested in Manchester United, for instance.
Monica: Perfect example.
Jim: And you set one up for Manchester United and you might actually further tune it that says, “All right, I only want game summaries and trades of players and nothing else, no gossip about the player’s, girlfriends or anything else.” And you define that as the hardcore sports Manchester United bubble and only things that meet that criteria enter the bubble. Is that about correct?
Monica: That is a perfect example. It matches my example of typing in World Cup and seeing what’s going on in the World Cup in whatever sport that is.
Jim: And of course, importantly as I was reading the paper, I said, “Some people are better than others at kind of understanding tools, online tools. And so bubbles may be something that we’d want a bunch of them floating around that people could subscribe to, preexisting bubbles or they could inherit bubbles and modify them.” I think I could see that as being something that would help accelerate the adoption of something like this, because probably most people would find creating a bubble from scratch to be either over their head or more work than they’d like to do.
Monica: Well, you can do a web search today and it’s the same effort. You just type in a topic and now you have a bubble for it. And by the way, if you click on somebody else’s message, you are inheriting the bubble of that message and so on. And you can always fine tune whatever bubble you’re in, you can always fine tune it and then the system and you close the window, the system will ask you, “Would you like to save this bubble in your playlist?” And you can do that.
Jim: Yeah, that would be good. And then with so searching, so let’s say if I search for Manchester United. Yeah, I might find some preexisting bubbles, right?
Monica: Yes.
Jim: That I could join. Or if I don’t, I could hit create a new bubble that starts out as just a blank Manchester United search and then I can tune it from there.
Monica: Yeah, if you want to be highly technical about this, come [inaudible 00:12:59] but bubbles will contain other bubbles. And for legal reasons, we may be forced in many jurisdictions to have an outermost bubble that says, “You cannot post about these topics.” And this of course is implemented with an AI that perfectly understand what these things are about. So it’s going to be a 100% reliable.
We will not post about these topics. And so that bubble is an external legal bubble if you will. And then inside of that bubble we can build a safe bubble that avoids things like spam and hate and the racism and all the stuff we don’t want to see. And people can just say, “I would like to subscribe to the standard, the safe bubble.” So now they have two bubbles already.
And so after that, everything they create is going to be inside of those bubbles. Now if you are for instance, curious about unsafe stuff so to speak, if you’re going here for instance, for porn or something like that, then you can unsubscribe to the safe bubble and just go directly to the legal one.
And if you’re law enforcement officer, and you are validated as such, you may be able to basically see what is the legal bubble it’s all filtering out? So in spite of having a system where you can say whatever you want, which means that you can do things like look for, I don’t know, credit court fraud or murder contracts or drug deals online, you can be sure that the law enforcement will be in those same bubbles. You don’t know who’s in the bubble you’re in because all it takes is for somebody to create something that intercepts messages that the bubble is about.
Jim: Now this actually may or may not solve a famously difficult question about the web, which is legal jurisdictions particularly about speech vary radically, right? In countries like the United States, you can say almost anything. There’s a small category what you can’t say, but it’s pretty small. In countries like Canada, there’s quite a bit more stuff you can’t say. In countries like Germany, there’s even more stuff you can’t say. In countries like Iran, there’s a long list of things you can’t legally say.
So let’s take our Manchester United example again. Is this embedded in a specific legal bubble or is somehow, can we magically have the legal bubble post-filter my Manchester United bubble so that it’s legal in my jurisdiction, but in a place like United States other than let’s say I’m in Iran, only in Iran legal stuff comes out, in the United States, US legal stuff comes out.
Monica: Yeah, this is at the level of implementation, which is I’m interested to be at this point, but it can clearly be done and it can be clearly enforced with an iron fist by artificial intelligence that understand exactly not only what the message is about but, your intent in the message.
And so if you want to solve a disagreement or strife in a meeting and people have said, “I don’t want any strife in here, be non-strife, non-argumentative [inaudible 00:15:58] bubble or something people can subscribe to, you basically will get a much cleaner field and it’s up to you to specify everything.
I mean except for the legal limit, everything is in the control of the end user, everything. Everything about which messages they see.
Jim: So it’s at that level, the bubble is personal.
Monica: Correct.
Jim: So it may start out as it’s an attractor, so it attracts things that match the content domain and then you add other filters to filter out what you don’t want to see now. Now, from a practical perspective, let’s imagine we start out with the Manchester United bubble and there’s threaded conversation going on, as there often is in these things. If one has a different filter than other people, one might see those threads kind of be somewhat incoherent, in that some messages would come through and others would not.
Monica: Yes. And there is a couple of remedies for that. I mean, [inaudible 00:16:58] people have to get used to this because that’s what a lot of the conversation is going to be. As you can individually basically filter out any opinion, you are going to have a more fractured conversation. However, once you’re in the thread mode, you’re looking at the real thread and everything is visible. So at this point you can add, like you just suggested. You can use this bubbles still as filters after the fact by basically seeing a message you don’t want to see in a thread, you swipe left on it.
If the finger is available and if the finger is not available, you can click on a mouse button and that will basically deprecate that message and you will not see any messages like that in the future.
In fact, every time this swipe left, they turn gray and there’s going to be collateral damage to other messages that also turn gray that you also didn’t want to see. You hit the refresh button and they go away. Conversely, if there’s something you really like, you can expand your bubble by right sweating on the sentence or a message and more messages concerning that topic will appear in that bubble.
So we have what we call tuning, which is a fine tune that you can use anywhere in the system and whatever bubble you’re currently looking at is modified to accommodate the swipe tuning and it’ll be saved in your plates if that’s where it came from.
Jim: Interesting. Now, I don’t want to get down to implementation details, but I’m going to give you an example of how a really old online system does this. It’s called The Well, it’s one of the oldest online systems in existence. It started in 1985. Probably 20% of the users are still using the command line version, believe it or not. And it has a very rudimentary filter called the Bozo Filter. Sometimes conversations get very contentious on The Well, and some people like contentious conversations and some don’t. And so people will Bozo Filter somebody in a given conference, which is kind of a equivalent of a forum, a collection of topics.
And I think it’s actually a pretty nice implementation, even though it was a very ad hoc decision based around the affordances of the software. It basically still has the post in it and it says Bozo Filtered and gives the time and the date and I think the author. And it’s not that hard to then retrieve the note if you wanted to read it.
So it might be interesting to have soft filters like that you keep the context there, you know the order of posts, but you don’t have to look at the post from somebody that you know don’t like or that has hate speech in it or something like that. But if you actually want to, you want to take the next volitional step, you could do so.
It’s actually a kind of simple but elegant solution to the problem of de-coherence around filtering out things in otherwise conversations that have an inherent sequencing to them.
Monica: I was briefly on The Well. So I know what it’s like and I have to say that many of the ideas that go into this are not my own. They’re basically gathered from all over the universe of social media and email and stuff. So there is, however, one interesting thing you can do in Bubble City, after you have basically you are in a bubble and you fine-tune it and it starts getting thin and you don’t know what’s going on, what to do? You can flip on the flag that shows you everything that the filter has removed and you can go in there and click, “Oh, this shouldn’t have been removed, this looks interesting.” You just click on it or right, swipe it. And suddenly your filter starts including those messages that were previously removed, possibly by accident.
Jim: Yeah, that’s kind of like going, that’s going through your spam directory and Gmail-
Monica: Exactly.
Jim: … every once in a while. Though, I got to say Gmail spam filter’s getting so good, I hardly ever do that anymore. But a system like this is probably going to be a little bit less precise initially and it would make sense to be able to review what’s been extracted and say, “Oh, sorry, that’s not what I wanted.” Or of course, in the future we could have our own personal bot do that for us.
Monica: Well, I mean Google is at the forefront with message understanding, I mean webpage understanding and everything. So there that stands to reason that they would use whatever AI they have to create a very good spam filter. And also they have the volume. They have the volume, they can see if the message goes out to hundred thousand people, they can start off thinking about whether this might be spam. So they have that leverage.
Jim: Okay, another question for you. As you say in your paper, there’s no like button and dislike button that’s publicly viewable as kind of an editorial comment, but instead you have a personal, “more like this, fewer like this”. Now, is that information about “more like this, fewer like this” used behind the scenes as an algorithm for helping to route and rate content?
Monica: Only in one place. And that is we create a special pseudo bubble, if you will, called Popular. And here we are looking at what people think is popular, and this is where the “more like this” information goes in. But the point to “more like this, and fewer like this” is that they are not going back to the author. The author cannot benefit from you liking or disliking them.
The only satisfaction they get is that people respond to their messages and that they have a reach count, which tells you how many people have seen it. And that’s basically all you get as a user.
Jim: I would suggest in addition to the Popular bubble, you also also have one called Shit Hole, for the stuff that has the highest negatives. I’d like to see that at least occasionally. See what people hate. That might actually be more interesting than Popular. I mean, Reddit has a popular thing and it’s usually boring as shit right? Idiot, public celebrities and crapola like that, right?
Monica: Well, we could have one which is basically messages favored by your personal AI and out of the whole set. And vice versa, if you want have a… We need a better marketing term.
Jim: Yeah, no, I think Shit Hole is perfect, actually. But now, the other thing though is I know you’re working hard here to depersonalize the filtering. However, I would say in my own experience, the imprimatur that comes from certain people is valuable, right?
Though, and this is kind of an interesting thing, I’ve used this example before. I can think of a particular person whose views on social systems on networks are really, really good. But their opinions about literary fiction totally suck, right? So it’d be nice to be able to, in some sense have a two-part filter which says person X, all right, they bump them up in status. But if it’s about literary fiction, don’t let it in.
Monica: Well, if you’ve just been creating a bubble with that person as a component, I mean their identity as part of the filter, which is how you follow people, and then adding to it the social context, we’ll give you exactly that because none of the other stuff will make it into that bubble. You don’t have to exclude anything. It’s exclusive by default.
Jim: Okay. Now let’s take another step, though. Look, imagine I have a community of people I follow on a topic and I want to build a bubble with 20 people, but I want to filter specifically some of them on specific categories but not others. Can I do that?
Monica: You would be an advanced user and you would’ve to use nested bubbles. That’s my opinion at this point. But we have four levels. This is in the next version of the paper. We have four levels of users, and the last level is called analyst. And the analyst level may have stronger user interface. And this is implied, this is aimed at, for instance, stock market analysts or law enforcement and emergency services and so on.
Jim: I got another possible use case for you-
Monica: Sure.
Jim: … which as I was reading the paper, it struck me that there was an ecological niche for curators to create good bubbles. Because creating a good bubbles could be a lot of work in maintaining it. And so Jim decides he’s going to curate the bubble about complexity science, let’s say. And I sit there and tweak it very carefully and add sources to it, put filters to reject bad stuff, et cetera. And then I might want to, I’d love to be able to license people to have access to my bubble. They can pay me a dollar a month to have access to my bubble. The money doesn’t… Because you make a good point about not wanting the flow of funds to go back to the content originators because it provides some perverse incentives on certain kinds of behavior.
But perhaps there ought to be an ecological niche for paid curators to build and take care of bubbles.
Monica: It’s possible. What happens is that certainly agencies such as law enforcement and others that are using these systems professionally to track by larger movements, they will have very carefully crafted bubbles. But most of the bubbles that people would be using in the system are kindly close to ephemeral, because things come and go, their interests was and wane. And so you typically have a playlist of maybe, I don’t know, 50, 100 bubbles, something like that. And you don’t really care about the last 60 of them because they were something you did earlier. And if somebody posted one of them, you can view it and say, “Nah, this is getting stale. I’m just killing the bubble.” And so you pop the bubble and it goes away.
Jim: At least within your own space. But I guess my point is that there are domains which are relatively well defined, let’s say in the sciences and in the early days of Twitter, there were people who spent a lot of time reading papers and retweeting the ones they liked. You don’t see so much of that anymore.
But if there were a ecological niche to get paid to do that, this might be a very good set of tools for doing that.
Monica: Well, it’s getting paid for it. I mean, people will use the system even if they have to pay for it. That’s part of my penny to post thing. That it actually costs money to post.
Jim: Yeah, I was going to talk about that next. So let’s go there.
Monica: We’re going to talk about… Yeah,
Jim: Yeah. Well why don’t we finish on the idea of curation as an occupation and then we’ll move to how people pay.
Monica: Right. Curation is extremely important and it’s extremely important in machine learning because we have to have curators create the corpora we use for our artificial intelligences.
So they’re going to be in high demand all over the place. Now, if you are a curator and you can curate your own bubble, you realize that every time you post it, when somebody clicks on your bubble, they inherit the whole thing. So it’s not something that you really can protect. But if it is well crafted, I mean it’ll spread. There is no way to protect a good bubble. How’s that?
Jim: Well, maybe there is. Suppose-
Monica: Maybe there is. It could be done on the outside, I don’t know, but it’s-
Jim: Well, here’s an extra feature I’m going to suggest to you. I didn’t see it in your writeup. Suppose I want to build a proprietary bubble and the bubble now has a feed and people can subscribe to the feed that come out of the bubble. They don’t subscribe to the bubble directly. Now they can add additional filters downstream. They can take the filter stream, connect it to their own bubble and then add other filtering. But the stream itself is an artifact of the curated bubble.
Monica: Oh, I don’t want to go there. It’s contrary to my principles in information here. It shouldn’t cost anything to read and so on. And also one of the points of system is that it federates all these other feeds into the system. This is how we bootstrap with it before we have any users. We basically federate mastodon and matrix and possibly Twitter if we can into the input feed. And you’re basically choosing on all those.
We might include blogs and substack and all kinds of other stuff there too. And the information feed is basically going to be very tech. It’s going to be a lot of stuff in the field and it’ll basically see within minutes of something appearing in any of these feeds. If it’s in your sphere of interest, you’ll see it in your bubble.
Jim: Okay. Well, I was going to ask you think about this idea of curation-
Monica: Yeah, I’ll think about it.
Jim: … with an output stream because it adds a new set of to topologies that it can emerge. I can see how it violates-
Monica: Right, right.
Jim: Yeah. So anyway, now let’s move on to the next question. Every venture capitalist always asks, “How are you going to make money off of this?” And you have a quite interesting and somewhat contrarian take on this. So how are you going to make money off this?
Monica: Yes. So getting an account is free, tuning your bubbles and defining your bubbles is free. It costs a penny to post. And the way you do that is you basically, you follow something. Imagine that you’re basically following the stream and you see an interesting topic and you click on a message and you want to reply to that message, which is what you do by default after you click on something, and you type in a little bit of a message. But the system has warned you several times already that before you can post, you have to have money in your purse. And so when you click the post button, it says, “Please add money to your purse using one of these credit cards on PayPal or whatnot.” And there’s a button there that says, “That’s outrageous.” And if you click that, the system pops up the window that says, “If your post isn’t worth a penny even to you, why would anybody want to read it?”
Jim: Yeah, I love that as a… That’s a great marketing line. And I actually read my… Initially, I said, “A penny can’t possibly be enough,” but I went up and looked at the numbers for how many tweets there are, and a penny is close, maybe it’s 2 cents, but it’s no more than 2 cents to support a system at the scale of Twitter-
Monica: Exactly.
Jim: … per tweet, which is quite amazing actually when you think about it. And per Moore’s law, that price will only go down over time. So if it’s 2 cents today, it’d be a penny in two years and half a cent in four years. And that’s actually quite remarkable.
On the other hand, we do know that doing large neural net stuff is not cheap, and that may cause the price to go up. In the early days it might be 5 cents. For instance, if we’re going to have heavy neural net processing. On the other hand, things like vector space look up is very inexpensive. So to the degree that we can process something once and then embed it into a high-dimensional vector space, the retrieval isn’t all that bad.
Monica: And that’s exactly what Bubble City does with a minor modification. We don’t use a vector space and well, it depends on which implementation mode. But yes, that’s exactly what Bubble City does.
Jim: Yep. So yeah. So a penny to 5 cents. And I love the marketing hook. Yeah, you pay to post, but if your shit ain’t worth 5 cents, don’t deposit it here. Right?
Monica: Exactly. I mean, we can play plenty of games with this. I have said for instance, that we can have the stamps will expire in a year, which guarantees $10 a year from every user, whether they post or not. If you have money in your purse, your account will not go away. But if you don’t have any money, it goes away if you’re not logging in for three months, et cetera, et cetera.
So putting money in your purse basically raises you from the status of lurker to the status of an author. And those are treated much better by the system.
Jim: All right, let’s move on to another one of the big design dimensions for systems of this sort. And that is whether posters need to be named real names or suit anonymous, i.e., one identity per human, but it doesn’t have to identify them as a human or can there be multiple identities per human?
Monica: And this is where I put on my marketing hat and said, “Okay, you can have as many accounts as you want. Each one is a nickname. Each one requires a different purse.” So you cannot basically do identity spamming economically because each identity that you create is going to cost you another 10 bucks, otherwise you can’t post on it. So that’s one strategy that we have for this.
Also, we say that everybody posts by the nickname only. So basically we don’t know who the person really is. We might do something like what Twitter has with these various blue markings of some kind. And specifically we might do what Elon has done about institutional stuff where basically if you are working for natural institutes of health, you can actually have an NIH badge on your thing, which is verified by the system.
And so we can do all of those kinds of tricks. But the main point is that once you post it, we have at least a credit card number to you. We don’t care who you are, but law enforcement might. And if they come and start [inaudible 00:33:53] us for who posted that thing, we can hand them a credit card. And that’s as far as we can go.
And by the way, we probably will outsource that to some other entities; log in and credit cards are painful.
Jim: Yeah, there’s lots of people that do that for a living. Now, what was I thinking about that here? Oh, yeah. Which is personally, I would very much like to be able to filter on the strength of people’s identities. I would like to be able to create a bubble, for instance, that says, “Real verified, real name only, double blue check or something.”
Monica: That never occurred to me, but that is an excellent idea. We’ll put it in, we’ll put it in.
Jim: Okay. Yeah, that, because I can see the argument for allowing the anonymous, but I can also see arguments for wanting to be able to create bubbles to keep all those people out. I’ve been involved in the creation of things like social media for, scary to think about it, 41 years now. And I will say as a general rule, non-real name content is worse than real name content. Not always the case, but it’s the way to bet.
Monica: Right. It comes down to, I mean, if you are behaving, if you’re not doing illegal stuff, nobody will ever want to know who you are. Nobody can identify how who you’re, because it takes a subroutine to get [inaudible 00:35:13]. I think that’s a fair way of doing this.
Jim: Okay. Now another issue that we all know, and I’ve written about this a little bit. I call the idea viscosity. Depending on how one defines a bubble, one could have far more messages than one would actually care to read. And you have a concept called Pacer. How would you deal with the problem of, let’s suppose I put up a bubble for Donald Trump. You’d get so much crapola coming into that. How would Pacer work to modulate that?
Monica: Well, first off is that basically Pacer is supposed to be used at things like the popular stream and it turns Bubble City into a television set where you’re basically passively sitting watching stuff flow by on the screen. And when you see something interesting, then you click on it and now you’re interacting. But up until then you can just sit there and watch the stream.
And the point is that all social media will throw away messages because if you have 2000 friends on Facebook, you can’t read what they post in a day in a week. So every message system filters away stuff that it for some reason thinks you shouldn’t see. And if you look at the very strong stream, a pet stream, like a Trump stream or even a popular stream, you don’t want to see everything. You want to see it sampled. And so you turn down the volume control. And volume control is basically speed dial. And you set it to, okay, starts out maybe one message per five seconds and you can go up to one message per second or something like that. But you basically said it to what you can constantly read on the screen.
And then you sit back, and you can of course adjust it if it’s too fast, as well.
Jim: You think there are people that actually do that. It seems like a bizarre thing to do. Sit there and watch tweets come in in realtime. I mean, I can’t imagine a more ridiculous way to spend one’s time.
Monica: Suppose they’re all about Manchester United.
Jim: I’m not that interested in it, but maybe some people are.
Monica: No, whatever. I mean if you are basically interested in epistemology and whatever, you create a bubble for epistemology and there may be way too much to watch there, but you can turn on the speed… Speed dial is one way to sample statistically rather than refining the bubble to something you can handle.
Jim: And so as I understood it, your Pacer literally just picks them at random, is that right?
Monica: It can pick at random. To be honest, we could cheat and we could use some popular ratings there too. But I think we will probably just pick at random, if we’re honest. Speaking of honesty, are you going to talk about the clickbait problem?
Jim: Yeah, let’s talk about it.
Monica: So clickbait is basically, it’s a nuisance. It’s everywhere. And it’s bait and switch basically. And what we want do is we do not have subjects lines on the messages, but you don’t see them in the scrolling bubble view. You only see summaries that the AIs have generated. These are honest summaries of the content.
And so you guarantee it pretty much if you click on a message, it’s going to be what the summary said it was about. And that gives you a level of comfort that you can basically just ignore stuff and people you usually do with these messages we don’t want to read.
Jim: You mean we won’t see TV stars from the 1970s, what they look like today?
Monica: Right? No, it’s [inaudible 00:38:57]. Yeah. Because the AI will recognize it as such.
Jim: Or we’ll summarize it into plain bland English. And if you want to hit click on it, go right ahead. But of course, obviously on social media today, because people are able to track the click-through rate on every title and they’re testing dozens or hundreds of titles to find the ones that are very strongest at pushing your buttons.
And so this is a good way to break that down. And you don’t have advertising directly. However, if people want to for a penny, they can post an ad if they’re so inclined, right?
Monica: Correct. And if they post an ad for Viagra, they will be heard from everybody who wants to see ads for Viagra and pretty much nobody else.
Jim: And smart people might say, “I only only want to look at Viagra if it’s less than a dollar a pill.” Maybe there’s some metadata that has pricing information in it. In which case then you could have smart and ethical advertising that people can parametrically filter their ads on metadata.
Monica: That’s walking in the reductionist direction. The thing to say is that the AI will recognize it as whatever it is.
Jim: Gotcha. Gotcha. All right. Let’s see what else we got here. Okay, playlists. We talked about swiping left and swiping right. Following popular, follow trusted streams. Okay, let’s talk a little bit about that. You can follow, you describe as trusted streams. What are those?
Monica: Reuters for instance? I mean-
Jim: Who decides? Who decides what’s a trusted stream?
Monica: Oh, they are named. You could follow Reuters by name. It’s something you can have a bubble for.
Jim: Okay. Oh yeah. And they could also follow Jim Rutt by name, right?
Monica: They could, yes.
Jim: Does Reuters have any different standing in the system than I do?
Monica: Nope. No people who are subscribing to it know what they’re getting in both cases.
Jim: So essentially all that that is, is a bubble with one or more name users as it’s input [inaudible 00:41:21]-
Monica: Yes. In some sense it gets closer to what you were suggesting earlier with a Jim Rutt channel coming into the system. But we haven’t even discussed or thought about making that something that end users could use. So basically it would be system level access to whatever streams that the system can subscribe to, would become available and they would possibly be available as named streams.
Jim: Okay, cool. Well, I’m going to go out and beyond your paper now a little bit, because as we’ve been talking about it, and as I extracted it from the paper, it seems principally around text messages of various sorts. Do you remember a product from a long time ago called Google Wave?
Monica: Yeah, maybe a little bit. I mean, it’s confusing. They have so many social media attempts. I don’t know which ones which anymore.
Jim: Yeah, yeah. Wave was interesting because you could create these containers that included messages. But the cool part about it, they also included artifacts. So you could actually build up documents. Multiple person edited documents. You could upload videos to them. And so these were smaller than bubbles, but they were kind of molecules, addressable molecules that people could subscribe to.
And so let’s say you’re talking about some very specific scientific domain. People could upload links. They could upload papers, they can upload videos, they can upload data sets even. Can you imagine moving the content of bubbles from just messages to artifacts or links?
Monica: That becomes basically an implementation level problem at some point. But I should mention that Google Wave, now that I think about it, I recognized it at the time as being somewhat Bubble City because they had these per topic bubbles, if you will. And it could be that they got this idea from reading my patents, which they owned.
Jim: And I went down that rabbit hole today because I remembered it when I was reading the paper. I said, “Ah, Google Wave, this is a bit like Google Wave.”
Monica: Sure yes.
Jim: And of course, I remember Google Wave came out and then disappeared and I wondered whatever happened to it. Turns out Google Wave was turned into an Apache open source project. And it still exists though apparently it never got much in the way of interest. There’s been one or two forks and there’s some kind of weak projects that are forks of Wave that still exist. But who knows, it might be worth looking at it as a way to bootstrap the development of your system.
Monica: It could well be. On the other hand, development systems like this are getting significantly simpler with things like ChatGPT programming for you. So I think I can implement this with a very small team and a lot of AI help.
Jim: Yeah, it’s probably quite… It is interesting. And I’m one of those people who will work on software for two or three weeks or four weeks or five weeks and then I won’t touch programming for a year. And I just recently went back to doing some programming. I built a GPT powered chat bot for the transcripts of my posts of my podcasts, actually.
Monica: There you go.
Jim: And, oh my God, does ChatGPT help a programmer who doesn’t program every day. It’s amazing. I mean, you can define even quite intricate little Python functions and it just writes them for you. It’ll create the template for a flask template for a simple website, just 10 words, press the button, it does it. It’s perfect. So yeah, it’s quite… Anyone who’s still Googling to try to find out programming tips, don’t do that ChatGPT is miles ahead on helping you be a more productive programmer, especially if you’re someone like myself who isn’t programming every day and whose fingers forget. My fingers forget Python syntax after six months or whatever. And then I go back to it, ChatGPT has really helped me be way more productive this time than I think I’ve ever been in my life, frankly.
So I think that’s a very good point that the cost of developing software has just suddenly made a discontinuous dropdown.
Monica: Correct? Yes. It turns out that basically junior programming skills are going to be less important going forward, but if you are a systems designer at the high level of systems architect, you can do a lot more by yourself than you used to be able to.
Jim: Yeah, I have tested GPT on big things and anything over a hundred, 200 lines of code, it kind of loses itself and just doesn’t do it. But of course that will improve. And so the skill today is more like a systems analyst skill. You decompose the problem into modules and then you describe the modules.
Monica: Exactly.
Jim: And ChatGPT is great at writing a 30 line function, for instance. It’s almost infallible.
Monica: And you don’t have to look up the API specification if you’re doing something on the cloud because it knows those, too.
Jim: Yeah, exactly. That’s the thing. I’ve already found. Even fairly obscure packages, it knows the APIs, which is quite impressive. And that will only get better. No doubt there’ll be products that are specialized for code, for instance, that will be better. The current Da Vinci 3 is pretty damn good, but I expect they’ll be better ones in the future. But they don’t have to be much better. They don’t have to be any better than they are today to be able to write pretty damn good 30 line functions in Python, I can tell you that.
Monica: Exactly. Yep, yep,
Jim: Yep. Which is quite remarkable. All right. So I think we’ve done a pretty good… Anything else that we have that we didn’t cover with respect to the aspects of the system?
Monica: I would like to at least touch upon the AI component. What you have been following this article, in this interview is basically version 2.0, and I have a new version 3.0 that is much heavier on the AI component in the system. Bubble City 2.0 document was written before ChatGPT came out. Okay, so it’s six months old and after ChatGPT came out, I had to redo the entire Bubble City document to basically re-target it for AI inclusion as a significant part of the system.
And at the very core of the system is this message routing algorithm that routes things to the users that have subscribed to them. And there’s three ways to do this. And the first one is to use the search engine in reverse, which is what I put in the patents for Google. The second one is to use something like the GPT APIs to classify the messages according to the end users. And that might be expensive. I mean, the web search is something like two cents, and the ChatGPT used to be 35 cents and now it’s cheaper.
Jim: So it’s about 2 cents now. And that’s for a very large context about a 4,000 token context. My little chat bot project uses big context as it’s trick.
Monica: Exactly. And then this is going to continue to get cheaper, but there might still be a economic incentive to find simpler algorithms. And I have one that I have created, which is basically a competitor to GPT library, which is called Organic Learning and Understanding Machine One. And they are described at my website in chapters eight and nine.
And if you go down the route of using Understanding Machine One instead of using GPT as the router, you can do it much, much cheaper because you own the entire version of the AI, which comes from the organic learning stuff.
Jim: Now have you actually built that model?
Monica: I have built. Yes, I have. The Understanding Machine One is available as a cloud service, and if you go to chapter nine, it describes the API and it gives you a link to GitHub to download some test code in Python and you can run that and the organic learner itself, and I’ve written 23 of them over the past 20 years since 2001, and two of them work, and that’s version 21 and 23. And 23 is currently the one that created the understanding that’s in Understanding Machine One.
And the cool thing about my organic learning system is that it is extremely effective, efficient. Because today to train up a ChatGPT, like AI, you would pay hundreds of thousands to millions of dollars in training fees, in compute costs and electricity. And my system can learn a significant amount of any language on the planet in five minutes running on the laptop without using the GPU. And that’s pretty significant.
So I basically intend to go for learning every language on the planet because it’s so easy. Learning medical English, medical Japanese, social media, Finnish, et cetera, and create a system for people to use those competencies to classify messages and in the future to create ChatGPT like interactions, which is what I’m working on right now.
Jim: Yeah, that’s very cool. Now, for the actual content attractor to the bubble, you don’t actually need a large language model. All you really need is a semantic space embedding calculator.
Monica: Correct. And so like I said, that’s what search engines are and that’s also in a very different way, but my Understanding Machine does. So it is a pure classifier, but it is much more competent than classifiers of old because it is based on my machine learning technology.
Jim: And of course there’s some other interesting things one could do. For instance, somebody I know just recently combined just a very simplistic two-dimensional space projection clustering algorithm on a hundred million tweets and then took the centroid of the clusters and searched out for the tweets and then ran them through GPT to summarize. And it produced an amazing topic map from a hundred million tweets.
And so there’s an example of essentially finding synthetic topics or synthetic bubbles in automated fashion.
Monica: Yes. I mean bubbles as such exist all over the place. They’re not just called that. So in some sense, all of these bubble technology is doable with all, most 20th century technology, but there’s no reason not to do it with at least a decent understanding machine or even a GPT like interface because it would be much better. I mean, one of the things I’m talking about is that messages with hidden intent where for instance, they contain discussion on some current event, but under the covers they are basically doing some persuasion of some kind, and then AI could detect that and basically flag the message as being slightly persuasive in this political direction.
And you could see those things already maybe as colors on the stream flowing by on your screen. You can see basically, for instance, what kind of messages you are getting in the process.
Jim: Yeah. Well this is extraordinarily interesting, and as I said, I’m talking to people almost every day about things in this space using our new technologies to help information essentially organize itself to our preferences, rather than to the preferences of the Facebook’s and the Twitter’s and the heaven forbid, Instagram’s of the world.
So when are we going to see something like this? You actively working on this project? Are you putting a team together? Where does this thing stand?
Monica: Well, I have fractions of teams together, but like I said, this can be probably be done by, I don’t know, a four person team in four months or something like that because it’s very much standard application. You have the federating step, you have a database like MongoDB, you have the Understanding Machine or the API to DPT, and then you have a user interface. And that’s basically not much work.
And I need a little bit of money to do this and I’m very poor at raising funds. So that’s what currently is holding me back.
Jim: Yeah. So anybody out there-
Monica: I’m actively working on it, actively working on it.
Jim: Anybody out there in Jim Rutt show land who would like to help Monica, get ahold of her, help her raise some money and realize the dream. Because as she says, at least the functioning prototype is not that expensive to build today. Quite remarkably.
Now, scaling it up to another matter and the pressurizing it to get to an actual network effect critical mass will turn out to be the hardest part. And that’s a different set of skills as well. So anyway, I’d like to really thank Monica Anderson for an extraordinarily interesting walk into her vision of Bubble City, information system of the future.
Monica: Thank you. This was fun.
Jim: It really was.