DeepSeek Neuron Stack
DeepSeek-V4.1-Flash-1017's 40 main layers, each with 384 routed experts of 2,304 neurons (35,389,440 in all) and one shared expert. Pick a task to light up the neurons the task map found selective for it.
Top neurons
|
|
|
|
The task map's top 100 lit neurons by S and top 100 by A, with their stored values. Grey rows are not lit for this task (high A, but S below 1 or not active). Click a row to fly to the neuron.
What lights up
A neuron lights up for a task when the task map's registered rule calls it selective: its energy for the task (how often its expert is picked, times how strongly the neuron fires) is at least twice the other tasks' average in both halves of the prompts (log2 score S ≥ 1), and it is at or above the layer's median energy for the task in both halves.
A group lights the neurons selective for at least one of its tasks. That union is larger than the task map's pooled group test, which was not saved neuron by neuron.
"All tasks" colours each neuron by how many of the 128 tasks it is selective for, and each expert by the average over its neurons.
The shared expert's neurons light up when S ≥ 1. The analysis saved their scores but not the median test, so they light up a little more readily than routed neurons.
Scores in the details are the weaker half's log2 ratio: S = 3 means about 8 times the other tasks' average. They follow which experts the router picks more than how strongly single neurons fire.
Each lit neuron carries its own values for the chosen task. S is its selectivity: the log2 ratio of its energy for the task (how often its expert is picked, times its mean squared activation) to the other tasks' average, the weaker of the two halves. A is its activation score with routing left out: the log2 ratio of its mean squared activation when its expert is picked for this task to the same for the other tasks, the weaker half; A is undefined where, in either half, the expert got fewer than 20 of the task's tokens or no other task gave it 20, and for the shared expert, where it was not computed. Share is the expert's share of the task's tokens. In the 3D view and on hover, S and A are rounded to 0.1; the top-neuron list shows the stored values.
Values are shown for lit neurons only. Every neuron's scores for every task would be 128 × 35.4 million, about 4.5 billion neuron-task pairs (9 billion numbers with S and A), too much for a page; the lit ones are about 252 million pairs, two bytes each, loaded one task at a time. Unlit neurons show their best and second task, as before.
Source: task map run tm1 of 13 November (2,048 answers, 128 tasks × 16 prompts), work log entries 39 to 42. This page shows saved results; it measures nothing new.