action verbs in how-to steps
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. 20 of Qwen3-0.6B. Its strongest snippets are mixed: they peak on different tokens. 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 tends to fire on the action word in a step-by-step instruction, such as 'grow', 'twist', 'work' or 'cut'. It does not clearly change the topic when turned up.
Written by Claude from this neuron's activations and steering outputs. Confidence: medium. 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.
-
flower. An easier way to propagate Clivia miniata is by division, which can be done at any time of year. If you notice offsets (small plants growing from<|endoftext|>Geography: Midway Atoll
-
skunk is a mammal, with a unique striped black and white coat and defense system: ( glands beneath the tail that produce an oily, sulfurous substance which
-
the patients bone marrow as opposed to bone (painfully) harvested from the patient. The mixture is used as a "scaffold" for the new bone to grow around/inside.⏎Vision correction can be achieved using a semi-permantent contact
-
positioned⏎so you can fit U-shaped wire over the flat end and into the holes.⏎8. Slip large wire into the folded canvas slot of the net and twist the⏎9. Cut the twisted wire to no longer than 2" in
-
turn off the power at the service panel for the circuit that you will be connecting into before beginning work. Double check to make sure the circuit is actually safe to work on by using a noncontact voltage tester. In many cases, more than one
-
scrap paper, the longer the better.⏎begin with one small line⏎duplicate the segment 3 times, as explained earlier⏎leave the last fold folded, and here’s where our loop begins. start by counting
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 been in the market of a new car and I have been getting it back from a previous owner. Now I have been in the current market of a car that has been been passed to her... I have been... been passed to
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
the reign of Suleiman the Magnificent. However, after continued unrest in the interior, the Turks evacuated the region in 1630, leaving it in the hands of the Shiite Zaidi imams of Sanaa.
-
destroyed by the Turkish occupation forces. After the Turks left around 1770 the monastery was a ruin. From these ruins a new church was build. In 18
-
resulting in a four-storey keep, a three-towered forebuilding and impregnable defensive walls.⏎From 1216-17 the castle was besieged by the French,
-
extended to 70 years. When the period of copyright protection has ended, the written document, musical composition, book, picture, or other creative work is in the public domain. This means that no one owns the copyright and everyone is free
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
- L14:1684 excites
- L16:2231 excites
- L18:2729 suppresses
- L19:1686 excites
- L11:1412 suppresses
- L16:1360 weakly excites
- L18:2195 weakly excites
- L14:2946 weakly excites
- L17:2673 weakly suppresses
- L14:1551 weakly excites
- L18:594 weakly excites
- L19:2064 weakly suppresses
- L13:2655 weakly excites
- L17:2977 weakly suppresses
- L13:2911 weakly excites
- L9:2467 weakly excites
- L16:2584 weakly excites
- L11:1299 weakly excites
- L16:2526 weakly suppresses
- L6:3045 weakly excites
- L19:796 weakly suppresses
- L6:2979 weakly suppresses
- L15:2975 weakly suppresses
- L18:1660 weakly excites
Feeds later layers
- L21:1723 suppresses
- L25:2021 suppresses
- L23:1734 weakly suppresses
- L21:1628 weakly excites
- L21:1022 weakly suppresses
- L21:375 weakly excites
- L26:791 weakly excites
- L22:1075 weakly excites
- L23:1747 weakly suppresses
- L21:1535 weakly suppresses 'Old' as part of a name or phrase
- L26:2047 weakly suppresses
- L21:981 weakly excites
- L23:2479 weakly suppresses
- L23:2980 weakly suppresses
- L26:1002 weakly suppresses
- L23:1894 weakly suppresses
- L25:1564 weakly suppresses word pieces before 'trans' or 'transfer'
- L23:1467 weakly excites
- L22:2776 weakly excites 'can' or 'could' before 'be found/read'
- L24:2293 weakly suppresses
- L22:1345 weakly excites
- L23:442 weakly suppresses
- L25:1723 weakly suppresses
- L21:1083 weakly suppresses
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
- ływ
- Indented
- 解决问题
- 삼
- ProductId
- cố
- 发力
- 对学生
- artykuł
- 米饭
Suppresses
- odelist
- donor
- 又被
- belonging
- 已被
- 遭
- donors
- redraw
- entieth
- blev