"at" in "at first" and "initial" words
A neuronOne of thousands of tiny dials inside each layer. It reads the text so far and produces a number, its activation, that nudges what the model writes next. in layerOne step in the stack the text passes through. Qwen3-0.6B has 28 layers, each with 3,072 neurons. 21 of Qwen3-0.6B. Its strongest snippets are very consistent: they mostly peak on the same token. It is featuredA neuron picked for the full treatment: steering outputs and a written explanation. About 3,000 of the 86,000 neurons are featured., so it has steering outputs.
What this neuron does
This neuron lights up on the word "at" in phrases like "at first" or "at the time", and it leans against the word "initial". When turned up, the model's writing got stuck repeating "initial", "initially" and "at first start", which matches what the neuron seems to push toward.
Written by Claude from this neuron's activations and steering outputs. Confidence: high. Think it is wrong? Flag it (opens a prefilled issue on GitHub).
Fires most on
The text windows where this neuron's activationThe number a neuron produces for a token. Large positive means it fired hard; negative means it pushed the other way. was highest. Amber marks positive activation; the outlined tokenA piece of text the model reads at once: often a word, sometimes part of a word or a punctuation mark. is the peak. Hover a token for its strength.
-
. We need to prepare an M-file which defines the equation and then call the subroutine (ode45) to do the integration. This method may seem mysterious at first because you call a subroutine, which in turn calls your M-file. Thus
-
as iron and calcium deficiency.⏎How can Celiac Disease be treated?⏎The only treatment required is a gluten free diet. It might seem difficult at first to enjoy a complete and balanced diet without gluten, but in fact, providing
-
different from a normal essay.” As students researched their four locations, they also had to locate them on a 21st century Google map. This was challenging at times as many locations no longer exist. Students had to figure out the most accurate
-
Countries⏎As you can see, the between online dating a girl through your country and dating a guy coming from another countries is great. It can be very confusing at first, but it’s wise to understand the different civilizations that exist just<|endoftext|>Why
-
online in both Greek and English, offers the visitor a fascinating tour in a simple and understandable way, utilizing modern audiovisual media. Even small fragments which seem insignificant at first glance, create an opportunity to reconstruct the facts,
-
matter what the circumstances were, and did what was right no matter what!⏎George Washington was indispensable in launching the U.S government. Although many opposed his political views at the time, Washington's leadership capabilities were unquestion<|endoftext|>Random changes in genes, rather
SteeringAdding a push along a neuron while the model writes, to see how its output changes. Turning it up or down.
Push this neuron up or down while the model writes, and watch what changes. The outputs were generated ahead of time, so nothing runs in your browser.
Here is something I have been thinking about lately: I have a 3D ZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZ
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
This tiny difference of starting values caused the equations to spit out a completely different set of answers. To mathematicians this effect is known as sensitive dependence on initial conditions and it is one of the hallmarks of mathematical chaos. It means that
-
Shareholders' equity, also referred to as net worth or owner's equity, represents the residual interest in the company's assets after deducting liabilities. It includes initial investments, retained earnings, and any additional capital contributed by shareholders. It
-
of understanding, completing 4 weeks of focused work, and then reassessing to check that progress has been made in key areas.⏎- Week 1: Initial assessment;⏎- Weeks 2 -5: One English, maths, and science
-
Generally, however, "people with disabilities" is the preferred initial reference.⏎Copywriters should portray people with disabilities as they would anyone else - with
Connections
Show on mapWhich neuronsOne of thousands of tiny dials inside each layer. It reads the text so far and produces a number, its activation, that nudges what the model writes next. in earlier layersOne step in the stack the text passes through. Qwen3-0.6B has 28 layers, each with 3,072 neurons. feed this one, and which later ones it feeds. Amber links excite, blue links suppress; longer bars are stronger.
Fed by earlier layers
- L19:457 strongly excites
- L20:2200 strongly excites
- L18:2514 strongly excites
- L17:2372 excites
- L20:1687 excites
- L16:885 suppresses
- L14:2632 excites
- L17:2719 excites
- L17:2441 excites
- L18:1899 excites
- L20:1798 suppresses
- L17:2922 suppresses
- L15:2918 suppresses
- L13:2977 excites
- L15:2881 suppresses
- L19:2013 excites
- L18:2532 excites
- L13:1143 excites
- L13:2367 suppresses 'at the beginning/start' time phrases
- L13:386 suppresses
- L11:1081 excites 'first' and 'inaugural' as starting points
- L16:1576 excites
- L11:1819 suppresses
- L19:2601 excites
Feeds later layers
- L26:1259 strongly excites
- L25:1396 strongly excites
- L24:1775 strongly excites
- L24:2617 strongly excites
- L27:910 excites
- L26:2398 excites
- L23:2062 excites
- L23:964 excites
- L23:938 excites
- L25:869 excites
- L23:3063 excites
- L25:1422 excites
- L22:1551 excites
- L22:2172 excites
- L23:1703 excites
- L23:743 excites
- L22:1518 excites
- L27:1004 weakly excites
- L22:2713 weakly excites
- L22:1505 weakly excites
- L23:958 weakly excites
- L25:2229 weakly excites
- L24:1815 weakly suppresses
- L23:289 weakly excites
Activation histogramA chart of how often a neuron takes each activation value.
How often this neuron fires at each strength. Almost always it does close to nothing (the tall bars in the middle); the snippets above are the rare moments it fired hard, amber for positive and blue for negative.
Logit lensReading which output tokens a neuron pushes toward or away from, straight off its output weights, before the later layers have their say.
The words this neuron nudges the model toward (amber) or away from (blue) when it fires. A hint, not a label: many neurons work through later layersOne step in the stack the text passes through. Qwen3-0.6B has 28 layers, each with 3,072 neurons. instead.
Promotes
- initial
- initial
- 初始
- Initial
- Initial
- initially
- 初期
- 一开始
- first
- →initial
Suppresses
- ┻
- olvable
- ugged
- тельного
- (condition
- ┴
- archy
- >|
- ently
- 況