Social Networks in Flux – The Shift from X to Bluesky
In recent days, we have witnessed a remarkable migration of users from X (formerly known as Twitter) to Bluesky. This trend is not merely a reaction to Elon Musk’s controversial stewardship of X, but also deeply intertwined with the outcome of the recent US presidential election. The re-election of Donald Trump and Musk’s prominent role in the new administration have sparked widespread debate, further accelerating this exodus.
Bluesky, originally developed as a project during Jack Dorsey’s tenure at Twitter, offers a decentralised social media platform granting users greater control over their data and communities. Within days of the election, millions of new users had joined Bluesky, sparking discussions on whether the platform could genuinely challenge X’s dominance. All indications suggest that this migration is far from over.
As an economist, I find the dynamics of social networks particularly fascinating. Unlike traditional markets, the value of a network depends not only on the quality of its product but also on the number of other users – the phenomenon we term network effects. At the same time, switching costs – the loss of followers, connections, and data – act as a powerful barrier that often locks users into existing networks, even when dissatisfaction sets in.
To explore these dynamics, I have developed a simulation that uses machine learning to model how users choose between two competing social networks. This model sheds light on why a platform like X can maintain stability for an extended period yet remain vulnerable to sudden shifts. In the following sections, I will explain how the model works and what it reveals about the future of social networks.
How Does the Model Work?
To uncover the dynamics of social networks, I’ve created an agent-based simulation where each user is represented by an AI agent.
These agents face a choice between two social networks – Twatter and BlueOcean – and their behaviour is guided by reinforcement learning, specifically Q-learning. This framework allows us to understand how individual decisions and collective patterns can lead to both stability and abrupt transitions.
Reinforcement Learning: Learning Through Experience
In this model, agents learn which platform offers the highest value through feedback from their experiences. This process, known as Q-learning, is a form of reinforcement learning where agents adjust their expectations of each platform’s value over time. Each agent tracks two Q-values – one for Twatter and one for BlueOcean – representing the expected reward of being on each platform.
When an agent selects a platform and receives a reward, the Q-value is updated using the following formula:
Here, α (the learning rate) determines how quickly the agent adjusts its expectations based on new experiences. This process mimics how individuals often choose social networks – by experimenting, evaluating, and adapting over time.
How Rewards Are Calculated
The rewards agents receive depend on three key factors:
- Network Effects: The value of a platform increases with its user base. The more users on a platform, the greater its value to each agent. To capture this, the model employs an exponential function, amplifying network effects as the platform reaches critical mass.
- Costs: Platforms impose a fixed cost on users, which could represent subscription fees, time spent, or mental resources.
- Switching Costs: When an agent switches platforms, they incur a penalty representing the loss of connections, history, and social capital. This penalty grows with the duration of their previous platform use.
The Decision-Making Process
Agents choose a platform based on their Q-values, typically selecting the one with the highest value. However, there’s also a small probability (ε) that they will randomly choose another platform. This ϵ-greedy strategy ensures agents continue to explore new options even after identifying a seemingly optimal platform.
Noise is also introduced to simulate unpredictable factors such as individual preferences or sudden changes in platform features, making the model more realistic and dynamic.
How the Simulation Unfolds
The simulation begins with an equal number of users on Twatter and BlueOcean, with agents randomly choosing a platform. Over time, agents learn which platform provides them with the most value, influenced by network effects, costs, and their own experiences.
At each time step, the model updates:
- Agents’ Q-values based on their experiences.
- The distribution of users between the two platforms.
- How network effects and switching costs impact the platforms’ attractiveness.
In the next section, I will share the results of the simulation and discuss how it reveals the mechanisms behind the stability and sudden collapse of social networks.
What Can We Learn from the Simulation?
The simulation offers profound insights into the mechanisms that govern social network dynamics, highlighting why platforms can enjoy prolonged stability yet experience dramatic shifts. These findings have not only theoretical implications but also practical lessons for established networks seeking to maintain their position or for challengers aiming to break through.
Network Effects: Strength and Vulnerability
One of the simulation’s key insights is how network effects – the increasing value of a platform as more users join – are critical to a social network’s dominance. In the simulation, Twatter maintains its position for over 200 time steps due to its critical mass of users, which makes the platform highly attractive to both existing and new users.
However, the same network effects that create strength also breed vulnerability. Once users begin leaving a platform, the same mechanism works in reverse: each departing user reduces the platform’s value for those who remain. This feedback loop explains why collapses often occur quickly and decisively, as seen in the simulation’s shift from Twatter to BlueOcean around period 300.
Switching Costs: A Double-Edged Sword
Switching costs play a crucial role in retaining users on a dominant platform. Early in the simulation, these costs slow the migration from Twatter, even as some users begin experimenting with BlueOcean. They act as a stabilising force, locking users into the incumbent platform despite frustrations.
However, when the value gap between the two platforms becomes too large, switching costs lose their deterrent effect. Users judge that the benefits of switching outweigh the costs, and once a critical mass begins migrating, the process accelerates.
This explains why dissatisfaction – as we see today with X/Twitter – poses a serious risk, even to a dominant player.
The Role of Randomness in Major Shifts
One of the most intriguing dynamics in the simulation is the role of randomness. Agents do not always act rationally but occasionally experiment with the other platform. These small, seemingly inconsequential experiments can, over time, trigger significant shifts. When some users discover that BlueOcean offers higher value, they set off a cascade that ultimately leads to Twatter’s collapse.
This dynamic mirrors reality: social networks often hinge on influencers, niche communities, or unexpected trends that catalyse larger migrations. Aspiring platforms can capitalise on this by targeting key individuals to ignite these cascades.
Stability and Fragility Go Hand in Hand
A critical takeaway from the simulation is that social networks can remain remarkably stable for long periods yet remain extremely fragile. Twatter dominates for over 200 time steps but loses its position in just a few once the tipping point is reached.
This demonstrates that even the most entrenched platforms cannot afford complacency. Small disruptions – whether user dissatisfaction or a competitor striking at the right moment – can escalate rapidly into seismic shifts.
Applying These Lessons
For established social networks, the solution lies in consistently delivering value to users. Frustrations, poor moderation, or a lack of innovation can erode trust, creating conditions ripe for migration.
Conversely, challengers like BlueOcean can break through by offering a superior experience and minimising switching costs – for instance, by helping users rebuild connections or transfer data from their previous platform.
Experimenting with the Model
An exciting aspect of this model is its interactivity, enabling you to explore the dynamics of social networks yourself.
The model, developed as an artifact in the Large Language Model Claude, uses reinforcement learning to simulate how users choose between Twatter and BlueOcean. You can tweak parameters and observe how changes influence the competition between platforms.
Have a look at the model here.
Explore Real-World Scenarios
The model is a powerful tool for examining real-world shifts in social networks. You can simulate scenarios where dissatisfaction with X leads to migration to Bluesky, or test how a new platform’s innovative features attract influencers and trigger a cascade of migrations.
By experimenting with the model, you can uncover how small changes in user behaviour or platform strategy can drive drastic shifts in social network dynamics.
In conclusion, this model provides a window into the future of social networks, offering a framework to decode the forces shaping their evolution. While social media often appears unpredictable, analyses like this help us understand how seemingly small decisions can lead to significant transformations.
Perhaps Elon Musk should take note.
PS you can find me on both Twitter/X (here) and Bluesky (here).
Picture created with fal.ai/FLUX.
———————
If you want to know more about my work on AI and data, then have a look at the website of PAICE — the AI and data consultancy I have co-founded.