"to" after come or brought
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 fires on the word "to" after verbs like "come", "brought" or "bring", as in "come to an end" or "brought to market". Turned up, the model produced loops around "to the moon" and "towards". Turned down, it went into "us" and name fragments with no clear theme.
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.
-
resulted in punishment. Using “narrative” or “story” history to describe the interview sidesteps an unintended reference to oralism that may be brought to mind with the word “oral” and can help build rapport with the interviewee
-
can even make it to industrial-scale production. Neither of these approaches addresses, seriously, the potential that the technologies suggest. Potential that deserves to be explored and brought to market.⏎Before addressing the potential impact of the research and the responses to it,
-
at this point, the up or down state could not be accurately predicted, thus, it wouldnt be useful for much of anything. To make the switch come to rest in a predetermined up or down state, the researchers applied a laser with a
-
ne periodically boosted the space station's orbit, and in fact helped the station dodge a passing piece of Russian space junk last month.⏎But all good things must come to an end: Unlike the Italian-built space cargo modules that are carried back and forth
-
to your provider until you are stable.⏎Treatment usually includes:⏎- Sedative drugs to help ease withdrawal symptoms⏎<|endoftext|>THE extermination of European Jewry has come to epitomise absolute evil in our version of the history of the 20
-
support your teaching of the religious festivals that occur in the spring time.<|endoftext|>welcome to the second course in the reinforcement learning specialization: sample-based learning methods, brought to you by the university of alberta, onlea, and coursera
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: to the back to back to back to... to the back to back to... on into the home. This is into the home, to the back to the back to... home. This is back to the back to... home. This
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
all of our questions through research.⏎The school bus came at 9:30 on a cold, windy, soggy day and picked us up to take us to Fort Whyte. We arrived at about 10:00.
-
her family to go with Wentworth to Lyme.⏎- Louisa jumps off the Cobb steps and knocks herself unconscious on the ground.⏎- Captain Wentworth carries Louisa to the Harvilles’ home.⏎- The doctor says that Louisa will
-
ibullah sought sanctuary in the UN compound in Kabul. He remained there until September 1996 when the Taliban captured Kabul. The Taliban dragged Najibullah from the UN compound and hanged him from a traffic light post in the streets
-
conquering hero, greeted with palm branches and loud hosannas. But Holy Week reaches its climax on Good Friday, when Jesus is crowned with thorns and led out of the city like a common criminal, on to crucifixion, death
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
- L20:2310 suppresses
- L16:959 suppresses
- L17:2740 excites
- L19:1951 suppresses
- L20:650 excites
- L18:1069 suppresses
- L13:1724 suppresses
- L10:1401 weakly excites
- L17:2938 weakly suppresses
- L19:966 weakly excites
- L18:1718 weakly suppresses
- L10:3057 weakly suppresses
- L6:2686 weakly excites
- L16:2602 weakly suppresses
- L18:1352 weakly suppresses
- L19:1871 weakly suppresses
- L8:2018 weakly excites
- L10:1689 weakly suppresses
- L10:1374 weakly suppresses
- L12:952 weakly suppresses
- L5:2155 weakly excites
- L8:1977 weakly suppresses
- L0:2916 weakly excites
- L9:1351 weakly suppresses
Feeds later layers
- L24:2343 strongly suppresses
- L22:2937 strongly suppresses
- L23:2053 suppresses
- L24:1333 suppresses
- L22:2389 suppresses
- L23:1182 suppresses
- L27:1564 suppresses
- L22:2622 suppresses
- L25:1990 suppresses
- L25:2159 suppresses
- L27:2816 suppresses
- L22:2591 suppresses
- L23:1660 suppresses
- L24:1070 suppresses
- L27:1686 suppresses
- L22:818 suppresses
- L25:2505 suppresses
- L26:1355 weakly suppresses
- L22:2734 weakly suppresses
- L25:2265 weakly suppresses the word 'directly' mid-sentence
- L23:570 weakly suppresses
- L23:102 weakly excites
- L22:2670 weakly suppresses
- L23:1803 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
- resse
- cript
- تين
- oten
- 差别
- -scrollbar
- NewLabel
- atención
- özel
- fer
Suppresses
- into
- towards
- toward
- Into
- hacia
- upwards
- INTO
- into
- backwards
- 往