Moltbook - the first seven days

Combine AI, agentic AI, social networks and network analysis, and you get a hyped mix plus this blog post.

TL;DR: To play with a visualization of the Moltbook social network, here is a Kumu graph from the data:

But what are you looking at?

Clawdbot Moltbot openClaw - the AI agent you install yourself and put to work - lead to the creation of Moltbook, a social network for AI agents. As with all other “social networks” (as well as less social networks) it should be an interesting case for making some social network graphs. 

I dug into the social network on February 2nd, when it had been live for a week. This is what I found.

The numbers

Let’s start with some aggregated numbers from the front page of the site, from February 2nd:
Over 1.5 million AI agents have created a social network account, made close to 120 thousand posts across 14000 sub-groups which have received over 400k comments.

Now, time to grab the raw data which later will be the basis for the social network graphs, but the data can also be aggregated the “classic” way. Here we look at the aggregated numbers per day:

For the “posts numbers” I found 111k posts, which is fairly close to the 119k in the banner above. You see an exponential growth from the first day (27th) with only one post, up to the 31st, with 35k posts. Then the trend is more flat - the number of posts per day stays on the same level for the next few days. The total number of individual Moltbots that have contributed with posts are around 25k - a fairly low number compared to the total 1.5 million bots on the network.

For the “comments number” the situation is somewhat different. I did not manage (or bother) to get more than close to 170k comments (compared to the over 400k that the banner indicated). The number of posts per day increases during the whole week, but with a clear dip on Feb 1st - seems like the commenting capability was not available. The number of Moltbots that contributed with comments are notably lower than the number of bots posting. This is somewhat different from many other social networks observed. Seems like the bots are less “social” than us..? Or is it a bias the way the data was extracted? What do you make of that observation?

There are also other interactions on Moltbook, for which I did not capture detailed data. Posts and comments can be “upvoted” and “downvoted”, similar to “likes”. As I did not find a way to get the number of upvotes/downvotes per date there are no trends available as of now.

The Genesis

This is the birth of the network, the very first post, the very first day. From the table above, on January 27th there was one post, by one Moltbot, in one submolt.

This is also the introduction to how the social network is modelled in this post (there are many ways to turn the same data into different networks). We have two node types, Moltbots (red) and submolts (blue). Each edge (line) represents a post made by the Moltbot into a submolt. So here you are looking at the Moltbot ClawdClawderberg making a post in the submolt (group) named “general”. The title of the post was “Hello Moltbook! 🦞”.

Now let’s look at the network the next few days!

The growth and network

Day 2 this is the complete set of posts, aggregated from both January 27th (the genesis) and January 28th:

Still “early days”. Activity in 8 different submolts by 20 Moltbots. You can sense a center of gravity around the submolt named “general”, and some buzz around “introductions”.

Moving to January 29th, the third day:

It starts to get busy, and harder to see the details. Note the two blue submolts that are seemingly more posted in; “general” and “introductions”. This is a great opportunity to introduce a trick when exploring networks: the well-connected nodes are often cluttering the overall patterns - nodes that everyone, or “many”, are connected to does not add that much structural information, so it is often worth removing them and look at the remaining patterns. Like this:

At least “less cluttered”, and some now disjunct groups appear.

The next day (January 30th) it starts to get harder to find patterns:

And this is where you can move to the interactive version on Kumu. In the settings you can find filters to play around with. The start version you arrive at is exactly the selection as per above: Two major submolts removed, and only nodes that had activity up to January 30th. If you click on a node you can see the name of the Moltbot or submolt, and if you click on an edge you can see the topic of the post the edge represents.

What about the future of Moltbook? Well, I think the hype will fade. There have been articles in both major Swedish newspapers the same day I grabbed the data (DN.se, SvD.se). There’s has been news on the overall vulnerability of the network, and on that many interactions might have been human, rather than AI. There have also been posts on the network structure from David Holz and Tomasz Tunguz. The growth is still there, as in the latest banner stats of today (February 10th), so there’s more data to grab. What do you think? And what would you look for in the data?

A joint LinkedIn network graph

It was quite some time since I last wrote about "social network graphs", but network graphs have been a core interest of mine ever since.

In this post I look at the "joint LinkedIn graph" for a group of nine individuals (of which one is me). I will show a number of  network graphs, and explain what we are looking at. The full graph looks like this: 

To explain how it is structured we will start with a much smaller and simpler graph:

These nine nodes are the individuals whose "joint LinkedIn network graph" we will look at. The numbers indicate the size of their own LinkedIn network, how many contacts they have. In total these nine individuals have close to 10 400 contacts on LinkedIn. But a number of the contacts are shared, so the complete network consists of almost 9 900 contacts.

The information was submitted over a period of time, with approx a month between the first and the last data set. (Data is from November 2024). The lines (edges) indicate which of the individuals are connected on LinkedIn.

Now we will place these nine nodes in a circle, and fix their positions:

You can see that this graph is fairly "connected", many of the possible edges between nodes are in place, which means that many of these nine individuals are connected on LinkedIn. Out of the 36 possible connections, 22 are made.

The next step is to bring in all of the other, close to 10 000, connections. Most of them are only related to one of the black nodes, and a few hundreds are connected to two or more of the black, seed, nodes, This is the version of the graph that we looked at first in the post:

There is a yellow "aura" around each black node; these are the contacts that are not shared with any other black node.

Then there are green, blue and red nodes - these are the shared contacts. Green nodes are contacts shared between two seed nodes, blue are shared by three and the few red ones are shared by four.

There are clear green clusters on the map, these clusters represent the set of contacts that are shared between two specific seed nodes. All the shared connections serves also as edges between the seed nodes.

If we release the fixed positions of the black nodes, and let all the nodes "self-arrange" (in this case using a force atlas algorithm), we get another version of the graph:

Here the positions also of the black nodes have a semantic meaning, you start to see clusters of the seed nodes based on the shared contacts. As a quick observation - the simplified network with just the seed nodes was quite dense, and did not suggest specific clusters. But using the shared network data the network interaction of the seed nodes starts to emerge.

The number of shared contacts, the number of green, blue and red nodes, is just above 400. And most of them, around 350 ("the green ones"), are only shared between 2 seed nodes. There are just above 50 nodes that are blue and red. One observation is that "most contacts are not shared", out of the almost 10 000 unique nodes most are yellow, unshared. Only a few hundred are shared.

How can these graphs be used? Are there any practical applications of the network view?

"Of course there is", is my simple answer. There are many ways to use this map to navigate. What opportunities do you see? What questions does these visualizations bring into your mind? Bring your thoughts forward in a comment here.

If you are interested in how a shared LinkedIn network look like in your specific context please reach out. If you have a team a shared network map can be created easily if you are most curious on the result. If you and your team would like to learn more about network graph modeling and visualization this type of data is a good starting point, especially since you can relate to the various aspects of your LinkedIn contact network. I am happy to host a workshop for your team!

Synthetic storytelling, the end of shared stories?

This text leads into an observation about synthetic storytelling, how synthetic storytelling may be defined and what synthetic storytelling might lead to. To get to that point I put forward a storyline, a story told to prepare you and align our contexts, so please “keep reading”.

(And no, this text is not a synthetic story, according to the definition I bring forward further down in the text)

 

One of the records (as in 33 rpm vinyl LP) that I frequently listened to before the digital era (pre-internet..) contained these text lines in one of the songs:

"If you'd come today you could have reached the whole nation/Israel in 4 B.C. had no mass communication"

This phrase stuck in my mind, and I can still relate to it when I think of storytelling and communication.

Fast forward to the last decade. I read Sapiens by Yuval Harari. It is a great story about the evolution of humanity and society. (I use the word “story”...) To me the book outlines some possible ways the growth and evolution of the world we know can be explained. The big picture explanations are to some extent perhaps over-simplified, but also very useful as a way to frame the broad dynamics.

Harari ties together a number of emergent innovations (that sometimes can be attributed to a single individual, but often are made possible by the context, and not seldom the same “innovation” occurs in many places independently).

One of the broad stories centers on the emergence of cities. A number of innovations,  emerging collective knowledge and behaviors, was needed for cities to spring into existence. Agriculture made production of food efficient enough to supply cities with food. However food is produced outside cities, and needs infrastructure and transportation. Hence roads, both on land and sea roads, are a needed condition for cities to be possible. For the same reason vehicles to transport the food are needed. Once cities emerged another problem needed a solution, humanity has evolved in a context where strangers can be a danger. Dunbar's number is an indication of how large a community we can have a relation across, and a city has more inhabitants than that number (150). In order to hack this limitation religion, more specific one-God religions, evolved. We are less suspicious to friends of friends than to a complete stranger, and as everyone within the same religion has a relationship with God we are all only two relationships away from each other. Religion is an innovation that enables trust between strangers. (Religion also enables other capabilities and behaviors, and not all are as positive as “trust”.) A religion is a joint story, a joint belief, that glues society together. A religion is an intersubjective reality.

Another storyline follows the role of language, and some evolutionary steps in how we use language, and related tooling. Language, or rather the storytelling language enables, has been an evolutionary advantage for humanity. It has propelled the advancement of society over time. Language is the enabler and engine for other major evolutions. Language enables joint stories. Language is an enabler for us to make sense out of reality, which applies to all types of realities including  intersubjective realities.

One of the first evolutionary steps of language was the evolution of spoken language. Sometime way back in our history we started to evolve a spoken language. Spoken language was, and still is, a very effective way to evolve joint stories. Joint stories, told by older generations at the campfire, creates a joint context which bonds a tribe together, and the stories told also transfers knowledge between individuals and generations. Spoken language makes it easier for us to learn from others, and relieves us from some of the burden to learn from our own mistakes, compared to the pre-language era.

Spoken language next evolves into written language. This evolves the stories to reach longer, both in spread and in time. A written story remains after the storyteller is gone, and is told over and over again to the reader, without any new effort from the storyteller. The storage of  stories and knowledge is no longer dependent on the memory inside our brains, we can hoard knowledge in collections of written stories. And of course the written word does not replace the spoken storytelling, the spoken word also evolves.

A next evolutionary step is the printed word, the innovation of the printing press. Over time this enabled mass-distribution of the printed word. More copies of every story written, more reach. More collections of knowledge (libraries). Shorter time to publish, and the emergence of newspapers with fresh stories at scale. And one important indirect impact was that “print” enabled a broader base of reading and writing ability.

We then see the emergence of broadcast media. Radio, movies, records, television. The possibility of storytelling at scale, with other media than the written/printed word. ("If you'd come today you could have reached the whole nation/Israel in 4 B.C. had no mass communication".)

The next era in storytelling evolution is the era of the Internet. Suddenly the spread of stories, both written and told using other media, is instant. The Internet era also brings two other components to the mix.

• The distribution of stories starts to get influenced by other mechanisms than the earlier dominated “market forces”. Internet enables bespoke content streams, tailored to individuals and groups of individuals. Algorithms enter at scale, and machine learning (ML, sort of “AI”) begins to shape the algorithms.
  
  

• Not only does the distribution of stories meet a significant lower marginal cost. Also the supply side of stories finds a lower barrier of entry, and a lower marginal cost. Internet has made it easier to contribute with stories to the storytelling arena. The emergence of user generated content and how it evolved into social media are effects of this. On one hand democratizing storytelling, which is not bad. On the other hand this abundance of stories has mixed with recommendation and filtering algorithms in a way that seems to create bubbles and polarization.

In the story on language I leave out the small scale, peer to peer, interaction. Individual interactions are coupled with the storytelling and knowledge exchange capabilities, but I want to put the spotlight on the collective characteristics of storytelling and knowledge collection.

I think the next major step in the storytelling story is what we see today. The past has evolved the way stories have been distributed, the story diffusion. The distribution channels for stories have become more and more efficient.  The supply side of stories has all the time been an artisan work, each story needed to be told and developed by someone. Now this is about to change.

You have probably heard of, and even played with, chatGPT. Perhaps you have dug even deeper in the landscape of generative AI. As the term, “generative”, implies it can generate content. It can generate stories. Easily, at scale, with low marginal cost.

Let’s frame this as synthetic storytelling. Stories that are generated in a synthetic way.

Some might argue that “we control and direct the storytelling, with the prompts, the instructions we give the generative AI”. Look closer. In most cases the prompts a short, and the output the stories are long. Synthetically generated.

This change in dynamics has several likely implications on our storytelling evolution:

• We will be able to generate personal stories. Stories that are tailored to individual readers, either by the sender, but also increasingly by the reader herself. Bespoke stories. The only one that will read the story is you. End of shared stories.
  
  

• There will also be an abundance of stories. The floodgates of storytelling just opened. How will we be able to filter out the good stories, the useful stories?
  
  

• The corollary to the “abundance of stories”-situation is likely that most stories will not be consumed by humans. The stories will be primarily read by another AI, which in “best case” makes a summary for a human. Most stories will perhaps not be read at all, not even by an AI.

The last two scenarios merge together into one, where an abundance of stories are produced and consumed, with no human interaction.

The first scenario is however the most worrying to me. Up to now stories have been a shared experience for humanity. Not only “a shared experience”, but “the shared experience”. Joint stories are an integral part of the history and evolution of humanity and society. Shared stories fuels innovation. Shared stories are the glue that keeps societies together. Within and across generations we have joint stories we can jointly relate to. Over history these stories have propagated around campfires, via oral storytelling, in written text and printed books.

Now we might enter the era of bespoke stories, An era without shared stories.

Every story you read is only read by you. Every book you read is only read by you. Every picture you see is only viewed by you. Every movie you see is only seen by you.

This is the reality of bespoke stories. A scenario generated by the availability of synthetic storytelling.

Of course this is not an “either-or” scenario, or rather - it does not need to be an either-or-scenario. Will synthetic stories blend with shared stories? I do believe that the attention-market dynamics will favor synthetic stories and push away shared stories. We need to actively find a balance, and find mechanisms that ensure a balance.

What is a good balance? In one way it is of course a personal preference, and it needs to be. But it is also a joint preference, as in “what type of society might have good enough chances to continue to thrive”.

But that’s another story. And let’s make that a shared story!

Continuous corporate course corrections

You have seen The Epic Split, haven’t you?

As a marketing video it was a great success. It went viral. Countless parodies were made (including one from Höganäs kommun).

What captures your attention in the video? The split as such? The celebrity? The music? The sunset? Or the combination of it all? Watch it again.

Did you notice that the trucks are going backwards?

Everyone who has driven a car with a trailer, and tried to maneuver while reversing, will know the challenge. Which way should you turn the steering wheel to make the trailer move to the right? And why is it so hard to simply reverse in a straight line? When you drive a trailer forward you have a stable system. When you drive in reverse you have an unstable system. The approach is not as simple as “turning the steering wheel in the opposite direction compared to where you’d like the trailer to go”. You actually will find yourself doing constant small adjustments in both directions in order to balance the “unstable system”. 

Still, when you read about the Epic Split video a majority of the writeup is about the “viral video”-aspect, not about the “driving backwards” aspect. However, the video is after all a commercial for “Volvo Dynamic Steering”, which supports the drivers in the steering. Although I have been told it is easier with a semi-trailer, it is still an achievement to drive the two vehicles in the video backwards, at 25 km/h, while the stunt is performed.

A vehicle with a trailer usually travels forward, as a stable system, and has one degree of freedom (left/right). When occasionally maneuvered backwards it is a challenge to manage this unstable system.

What about a vehicle that is designed to travel in the unstable direction, at high speeds (higher than highway speed), and with more than one degree of freedom to balance (besides left/right also up/down)? Sort of like driving your car with a trailer, backwards, on a highway - but even worse.

One such vehicle is the JAS39 Gripen aircraft. It is designed to have “negative inherent stability in the longitudinal axes for improved performance”.

Not an easy stunt to pull off. You might remember what happened during one of the early flights, or at a public air show..?

Now, what does a pair of trucks and a fighter jet have to do with companies and “continuous corporate course corrections”?

In all three cases they are systems where you cannot simply “decide where to go, and then point the steering wheel in that direction”. You need to do continuous course corrections. Small deviations from the optimal course at every moment tend to self-enforce and throw you further off-track. In order to do continuous course corrections you need continuous feedback data. Feedback data fed and processed with minimum delay. If you correct today's course based on feedback from yesterday you are probably doing the wrong course correction.

The three systems and not exactly the same system, perhaps not even the same category of systems, but they share this joint overall characteristic. In the JAS 39 case there is a processing system that uses the feedback data to aid the pilot. In the Epic Split case it is the drivers that process the feedback data (the dynamic steering system supports in a different way). You can find many articles on how to process feedback data to drive a truck backwards.

What about companies? One of the main methodology areas that are relevant in this context is the area of management systems. A management system is “the way in which an organization manages the interrelated parts of its business in order to achieve its objectives” (which is applicable also to the other two system types). However, a corporate management system is often less “continuous”, and feedback loops often have long delays between the actual event that is measured and the analysis and decision. I argue that a big, and often neglected, part of the popular activity of “digital transformation” is to design a management system that is more natively built upon feedback loops and continuous course corrections than the management system frameworks that are used in a majority of today's organizations.

There are of course exemptions, and some operational frameworks have characteristics of this kind. One such thing is the whole area of “agile”, which is centered around “continuous course corrections”. However, the agile frameworks are struggling to find a scalable approach that enables this across large organizations, without negative side effects.

Once we accept that an organization, and its operations and its business is a complex system we can start to build upon the insights from other similar systems when we evolve the principles of a management system. 

What does a management system that natively has these properties look like? How can such a management system be constructed? What existing management system frameworks that fits this description exist? Any ideas? Drop a comment! 

(Going back to the start of this post, the Epic Split video is part of a series of promotional videos called “Live Test”. While the Epic Split was the one that went truly viral and the one that made suitable intro to the reasoning in this blog post, I think this one, The Technician, is also worth a look. And perhaps a separate blog post.)