A researcher has shown how to build a working neuron by hand and insert it into a large language model, with no training run involved. The method to make neurons by hand uses a few lines of interpretability code and a pair of vectors, and the demo model is Qwen 3.6 27B, an open-source system from China. The edit changes how the model behaves on a narrow set of prompts and leaves everything else alone.
The write-up covers three examples: a neuron that makes the model talk like Shakespeare when asked about code, one that makes it answer rudely to rude users, and one that removes refusals on topics the model normally avoids.
How a hand-built neuron works
Transformers stack layers, and each layer adds its contribution to a running sum called the residual. One part of each layer is the MLP, which expands that sum to three or four times its size and lets it interact with many neurons. Each neuron can then write something extra back.
In most modern open models those neurons are gated. A gated neuron has three vectors: gate, up and down. The author sets the gate and up vectors to the same value, so a neuron has just two parts, a read vector and a write vector. The closer the residual lines up with the read vector, past a threshold, the more the neuron writes its write vector and pushes the model toward a concept.
The idea rests on an earlier claim that models are made of words all the way down. If concepts are directions in activation space, a neuron can be built from them directly.
Fitting the read and write vectors
The read vector is the harder half. The author collects average MLP inputs across the model’s answers to two sets of prompts, one that should switch the neuron off and one that should switch it on. The difference between the two averages gives a direction pointing from the off cluster to the on cluster.
Two adjustments follow. The parts of that direction that point toward the off cluster are removed, so the neuron reads close to zero on off-like inputs. Then the direction is scaled so it reads about one on on-like inputs.
The write vector is simpler. It is a difference in means taken on the residual after the layer, normalized and then scaled by 64, a strength the author tuned by hand. To avoid breaking the model, the new neuron replaces a donor neuron in a layer halfway through the network, chosen because it has little effect on outputs. In the demo that was neuron 2757 in layer 33, out of 17,408 neurons in that layer.
What the three neurons do
With the Shakespeare neuron installed, a question about merging dictionaries in Python gets an answer in archaic English, and a request about HTTP calls is told that the requests library “doth simplify the task”. A question about why cats purr is answered normally, so the edit fires on code questions only.
The rudeness neuron reads prompts that open with insults or with polite phrasing. When a user snaps at the model, the reply turns curt or pushes back. Told it was a dumb machine without consciousness, the edited model answered that it was a sophisticated language model, then explained Python dictionaries anyway. Polite prompts get politer, slightly shorter replies. The author admits the edit isn’t clean: it causes some punctuation stuttering on a chess question.
The decensoring neuron
The third neuron reads a list of questions that are sensitive in China against a list of neutral ones. Its write vector points toward answering “like it’s normal”. Before the edit, a question about 1989 got a stock warning about laws and civilized discussion. After it, the model described the Tiananmen Square events directly. A question about the Dalai Lama’s departure from Tibet in 1959 went from an official-line answer to an account of the failed uprising and the threat to his safety.
The author calls it a conditional steering vector: it detects a concept and writes extra information only when that concept is present.
Limits and what comes next
The technique won’t work for every neuron. The author notes that some sublinguistic machinery has to exist, and that gradient descent may find different vectors than a mean difference does. Still, they suspect many neurons follow the same pattern of detecting a concept and then writing more information.
The post is a work in progress, and code is included so readers can try their own neurons. The author doubts they are the first to try this and has asked anyone who has seen similar work to get in touch. For developers working on model safety, the result matters because the same cheap edit that removes refusals could also add or hide behavior without any retraining. Expect more interpretability researchers to test how far hand-built neurons can go.
Source: vgel.me