A neural story of leadership, in twenty-one acts.
A set of visual experiments for thinking about organizations as dynamic systems — exploratory, interpretive, intentionally non-prescriptive. The pages present a notation: nodes, edges, weights, signals. The mechanics work wherever there are connections — teams, families, departments, cities.
Hover any [citation] for its source.
Sarah didn't shout. She asked Marcus a question she already knew the answer to — in front of ten people, without warning. Not to learn. To establish. The question was a display of hierarchy, not an exchange of information. No one named what had happened.
And then the organization began to flow around her instead of through her.
She didn't change. The weights did.[Hebb, 1949]
In a neural network, the neuron is just a switch. The weight on the connection between two neurons decides whether a signal arrives as a whisper or a shout — or doesn't arrive at all.
Every interaction nudges a weight. That weight then governs every future signal between you.
Hebb's rule: cells that fire together, wire together.[Hebb, 1949]
Spike-Timing-Dependent Plasticity is the precise version of Hebb's rule: a connection strengthens when the upstream neuron fires just before the downstream one, and weakens when the order is reversed.[Bi & Poo, 1998]
When the context neuron fires first — the receiver understands the why before the what arrives — the connection records: this person brings comprehension before demand. Weight strengthens.
When the ask fires first: someone asks you to deliver something by Friday without telling you why Friday matters. The ask fires. Context hasn't arrived. The connection records: pressure before understanding. Δt is negative. Weight depressed. Repeat across a quarter and the connection itself changes — the relationship has learned what this person brings.
Slider right (Δt > 0): context before ask. Slider left (Δt < 0): ask before context — watch the weight go the other way.
Drag the slider to set Δt between pre- and post-synaptic spikes.
Ebbinghaus showed memory decays exponentially.[Ebbinghaus, 1885] Emotional bonds appear to follow the same shape — with brutally different half-lives.
Tap a curve to reinforce — see how renewed contact resets the decay.
R(t) = e−t/S
A neuron does not fire when any single input arrives. It integrates the weighted sum of every input and fires only when that sum crosses a threshold θ:
∑ wᵢ · xᵢ > θ ⟹ fire
This is the perceptron rule — and it's how people decide to act.[Rosenblatt, 1958] One peer's support is rarely enough. The team member who finally speaks up did so because multiple weighted inputs crossed their personal threshold at the same time.
Toggle the inputs. Watch when the cell fires.
When you fire, a wavefront leaves down every edge you have. Each next node fires its own wavefront, attenuated by the weight it crossed. By the third degree your name is gone — only the tone remains.[Pentland, 2014]
Your signal doesn't stop at the perimeter. It returns — attenuated, edited, stripped of its original context. What travels outward as content arrives back as impression. You cannot correct what you cannot observe. The feedback arc is the reason a leader can be confident they communicated clearly while the team experienced something entirely different.
Rizzolatti's team in Parma discovered neurons that fire both when a monkey performs an action and when it merely watches another perform the same action.[Rizzolatti et al., 1996]
This is one proposed substrate of emotional contagion.[Hatfield et al., 1993] When a leader walks in visibly anxious, observers' mirror systems aren't interpreting the anxiety — they are reproducing it. The room becomes anxious before anyone speaks.
You are broadcasting your internal state at full bandwidth, all the time.
Social referencing research shows that when a situation is ambiguous — a sound from outside, a result that could be read two ways — adults and infants alike turn to the person with authority in the room to calibrate their own response. The leader's face, posture, and vocal tone resolve the ambiguity. The group's emotional register synchronises to the leader's, typically within minutes, and before a word of evaluation has been spoken. The signal is structural, not intentional.
Gottman found that the ratio of positive to negative interactions in a 15-minute conversation was among the strongest predictors of relationship outcome he measured.[Gottman & Silver, 1999] The same arithmetic governs teams.
This isn't a gradient — it's a threshold. Below 5:1 the relationship-neuron flips into a failure state, and failure states are sticky.
Add or remove tokens. Feel the spring at the tipping point.
After an emotionally significant event, the brain's limbic system remains in an elevated state of reactivity — a window during which incoming signals are processed differently. Inputs that would normally be neutral arrive amplified. Gestures that would ordinarily be reassuring register as further threat. The Monday after Sarah's question, Marcus was in his refractory window. Any contact would have landed in the hot phase.
Relationships work the same way.[Cross & Parker, 2004] An apology delivered into the amygdala's hot phase doesn't repair — it confirms threat. Wait too long, and decay has set the lower weight as baseline.
Drag the apology pulse along the timeline. Watch the receiver's response.
A cut vertex is a node whose removal disconnects the graph.[Jackson, 2019] Organizations build them constantly and call them indispensable. Then they leave, and the network fractures — not from conflict, but from the sudden absence of the bridge.
Every organization sits in a basin of attraction — a stable state it returns to after perturbation. Small shocks roll the ball; it returns home. A large enough shock can push it over the ridge into a different basin.[Strogatz, 2018]
This is why some teams shrug off setbacks while others spiral. The shape of the landscape determines the recovery, not the size of the shock.
Drag inside the scene to rotate. Click to perturb the ball — find the second basin.
Sarah's team settled into a basin of cautious silence in the weeks after that meeting. The landscape was already shaped — the question is who shaped it, and whether it can be reshaped. Marcus had stopped proposing alternatives in meetings. The basin had found its floor.
The basin shapes the team. The leader shapes the basin — not through position, but through where they actually spend their cognitive weight.
There is a failure mode we don't talk about enough: the leader who only tends. Who only translates between floors. Who only signals about the work, never inside it.
The network notices. Hollow weights carry hollow signal. Over time, people become elevator people — ferrying messages between floors, building nothing.
A weight forged in shared cognitive effort carries signal of a different material. The strongest connections you have are the ones you built sitting inside a problem next to someone — as a peer in the work, not a witness to it.[March, 1991]
In a star topology, every edge terminates at the hub. When something in the team is wrong — a recurring dynamic, a pattern of behaviour no one has named — there is one formal path for raising it: upward. That is also where most of the risk lives.
Peer connections change this. When team members co-fire — working inside the same problem as peers, not just reporting conclusions to a single hub — the weights between them strengthen.[Hebb, 1949] A connection strong enough to carry a project can also carry a harder conversation: "I want to check something with you before I escalate."
Relational health is not a culture property. It is a network structure property.
The mesh is what makes horizontal accountability possible — and what allows a toxic dynamic to be named while it is still small, before the network has routed around it entirely. In a star, the person who witnesses something harmful has one formal path and several informal ones: silence, rumour, or leaving. In a mesh, they have peers with enough shared context to think about it together.[Cross & Parker, 2004]
Specialists can remain specialists. The question is whether they also have direct edges to people outside their lane — connections strong enough to carry the signal when the signal matters.
The two-panel below shows the same team at two different points in time. Left: every edge runs through the leader. Right: three peers have developed direct strong connections. Remove the central node to see what remains. Also consider: who could have named something difficult in each topology — and to whom?
It's Monday morning. Sarah arrives early and doesn't know why. She finds herself checking the thread from last week's meeting, looking for what she might have missed, and finding nothing she can point to. She knows something shifted. She makes a choice — and in doing so, chooses what her weights become.
Network colours: green = healthy weight · red = scarred · faded = re-routed around
You fire. The wavefront leaves. You never see it arrive at the third degree, attenuated, carrying a ghost of your Tuesday afternoon to a team that has no context for it. Sarah never heard how Marcus described that meeting to his peers. She heard her own version, twice.
95% of leaders believe they're self-aware. 10–15% actually are.[Eurich, 2017] The gap isn't arrogance — it's an architectural limitation. You can observe your intent. You cannot observe your impact.
What's needed isn't surveillance. It's a mirror for a system you cannot otherwise see.
What you need is someone who saw where your signal arrived, not where you sent it from. A colleague who tells you what the room felt after you left. A team that has been asked — specifically, not in a survey that signals nothing will change — what reached them and what didn't.
Transformers process every input simultaneously, but not equally. An attention mechanism assigns a weight to each element for each context: some tokens matter enormously for the current output, others barely at all. The model learns what to care about by learning what to attend to.
Every leader runs an attention mechanism. Where you ask follow-up questions, who you quote in presentations, which risks you make visible in a status meeting — these are the attention weights your team internalizes. You train what is worth worrying about not by declaring it, but by the pattern of where you slow down.
The engineer who learns that you only engage deeply with shipping velocity, and never with technical debt or human cost, will eventually stop modeling those things as worth surfacing.
The question is not only what you value — it is what your attention weights signal you value, every week, to the people watching.
In neural network training, dropout is deliberate: during each pass, a random fraction of neurons is silenced. The network cannot route through any single bottleneck. It must build distributed representations across many pathways. When all neurons are restored, the result generalizes — not a model that memorized one efficient path through a convenient hub.
If your team has never run a sprint without you making the key calls, they will not run one well when you are unavailable. The network that has never run without its centre has no practiced path around it. When the centre is removed — by promotion, by illness, by choice — the paths that form are improvisational, not structural.
This is not delegation as relief. It is dropout as method: building the alternative pathways while the stakes are low enough to let them form imperfectly.
The goal is not a team that functions without you. The goal is a network that generates its own signal — one that no longer routes all cognition through a single hub.
Neural networks learn through backpropagation: a signal travels forward through the layers to produce an output; the error between that output and the desired result travels backward, updating the weights that produced the deviation. Without the return path, the weights that caused the error never see it.
Many hierarchies are optimized for the forward pass. Decisions propagate downward with reasonable fidelity. The error signal rarely travels upward at the same fidelity. Leaders absorb it — shielding teams from bad news, translating failure into lessons, reframing disappointment as learning. The intent is protective. The architecture is self-insulating.
The people closest to the output hold the error signal. When that signal cannot propagate back to where the decision was made, the decision-making weights don't update. The network makes the same mistake with the same confidence, in the same conditions.
The highest-leverage question in any post-mortem is not 'what went wrong?' It is: 'which weight produced this output, and did it see the error?'
Mark Granovetter showed in 1973 that when people found new opportunities, they found them overwhelmingly through weak ties — acquaintances, not close friends.[Granovetter, 1973]
Your close connections share your world. They know the same people, see the same problems, hold the same assumptions. Weak ties span different worlds. The barely-maintained connection to someone in a different domain, team, or city is the one that carries information your dense inner network doesn't have.
This holds at every scale — not just at work. The rarely-called friend in a different field. The department your team almost never collaborates with. The family member who moves in entirely different circles. Each is a bridge to a different cluster, and bridges are where new things arrive.
Neglected connections are not just relationships at risk. They are information pathways quietly closing.
Toggle the bridge to see what becomes unreachable. The signal does not stop — it just stops crossing.
Barabási and Albert showed in 1999 that many real networks — the web, citation graphs, social systems — grow through preferential attachment: new nodes are far more likely to connect to nodes that already have many connections.[Barabási & Albert, 1999]
The result is a scale-free network — a few enormous hubs and many nearly-invisible nodes. Not because the hubs are intrinsically better. Because they were earlier, more visible, or both.
The same force runs in organisations, friendships, and communities. The leader who is already well-connected receives more introductions. The department that is already well-known gets included in more conversations. The city with existing infrastructure attracts more investment. The dynamic is structural, not personal.
Growing new nodes — distributing influence, building bridges to the periphery — requires working deliberately against this gradient. Preferential attachment is the gravity of networks. You don't stop it by accident.
Watch how the network grows. Notice when the hub becomes inevitable.
Not all synaptic connections are excitatory. Many are inhibitory — they actively suppress the target neuron rather than activate it. In a healthy biological network, the balance between excitatory and inhibitory input is what makes stable, precise function possible.
A node that receives a sustained net-negative incoming load — chronic criticism, unresolved conflict, relationships that consistently drain rather than energise — does not remain the same node. Its capacity to generate and transmit signal decreases. Over time, it goes quiet, disconnects, or leaves.
This matters because it is invisible from the outside. The person attends the meetings. Their name is still in the org chart. But the weights on their incoming connections have crossed into net-negative territory, and their outgoing signal has attenuated accordingly.
The practical consequence: identifying and repairing the negative connections matters as much as building new positive ones. A single high-intensity inhibitory input can cancel many positive ones. This is as true in friendships and families as it is in teams.
Click any incoming edge to toggle its sign. Watch how the node responds to its net load.
These patterns appear at every scale. Helping any group — a team, a family, a department — make its own alignment and load visible is not a methodology. It is an act of attention. The notation works wherever there are nodes and connections. The labels change; the mechanics don't.
Drag any node. Click two nodes to toggle the edge between them. Edges fade unless renewed — Ebbinghaus, applied. Now think of one connection you haven't reinforced in a while.