years and numbers before the digits
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. 25 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 small word just before a year or other number, such as the space before '1712' in 'in 1712' or before '12th Century'. When it was turned up, the text did not clearly change in any readable theme, so its job is hard to pin down from the outputs.
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.
-
is not just a hugely significant place of worship, but a gigantic repository of architectural and artistic splendour. There has been a church on this site since around 627 AD, when a wooden structure was erected over some part of the
-
sale.⏎The term originated with the cistercian monks at Eberbach Abbey in Rheingau, where the first recorded use of the term occurred in 1712. The abbey’s best wines were set aside to be
-
clues as to what life was like in medieval Edinburgh,” he said.⏎Excavation work has already revealed the precise location of the monastery, which was founded in 1230 by King Alexander II, the ruler of Scotland from 1
-
arrived in Scotland after the Norman conquest of England in 1066. There is a record of a David de Boivil witnessing a charter as early as 1164.<|endoftext|>For the questions below, use the most efficient Excel
-
the use of black and thick wooden. Most Gothic glass windows have large, pointed, Gothic-styled windows.⏎Gothic architecture has been around since typically the 12th Century. This was founded upon the idea that architecture need to be
-
we read:⏎Ebionites. This designation was at first like 'Nazarenes,' a common name for all Christians, as Epiphanius (d. 403) testifies (Adv. Har. xxix.1
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: how to align the learning stage with the stage of learning and align learning stages properly with learning stages of the person's growing stage phase, what stage stage stage stage stage stage stage stage stage stage stage stage stage stage stage stage stage stage stage stage
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
8⏎Ben sold 830 newspapers. His brother James sold 177 fewer newspapers than Ben.⏎(a) How many newspapers did James sell?⏎(b) How many newspapers did they sell altogether?⏎In early 2-step problems,
-
c) is called programmable array logic , first introduced by Monolithic Memories as the PAL series of devices.⏎Because the arrangement of Figure 5.11 (c) is very similar to a ROM, we sometimes call a horizontal product-term
-
the tables representing your daily salary for your first two weeks at work, assuming you take the job.⏎b. Write an equation to model the growth of your salary.⏎c. How much will you
-
construct the product terms) that feed OR gates (which form the sums of the product terms). We can simplify this representation still further ( Figure 5.11 c), by drawing the input lines to a multiple-input AND gate as if they
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
- L24:485 suppresses
- L23:333 suppresses
- L24:658 excites
- L23:2220 excites
- L23:2334 excites
- L22:1984 suppresses unclear: word before a count or quantity
- L6:1701 weakly suppresses
- L22:672 weakly suppresses
- L16:3071 weakly suppresses
- L7:1017 weakly excites
- L10:1304 weakly suppresses
- L24:2080 weakly excites
- L23:2364 weakly excites
- L1:984 weakly suppresses
- L10:858 weakly excites digit 1 in COVID-19 and 1941
- L21:2932 weakly suppresses
- L9:2422 weakly suppresses
- L7:466 weakly excites
- L6:1775 weakly suppresses
- L22:2 weakly excites
- L2:2910 weakly suppresses
- L18:1315 weakly suppresses
- L9:2480 weakly suppresses
- L0:1012 weakly suppresses "of" after nouns like eve, family, modes
Feeds later layers
- L26:1055 strongly excites
- L26:546 strongly excites
- L27:495 strongly excites
- L26:297 strongly suppresses
- L26:1427 excites
- L26:519 excites
- L27:779 weakly excites
- L27:629 weakly excites
- L27:18 weakly suppresses
- L27:668 weakly excites
- L26:1353 weakly excites
- L26:2543 weakly suppresses
- L26:1942 weakly suppresses
- L27:3012 weakly suppresses
- L26:2859 weakly suppresses
- L27:2710 weakly suppresses
- L26:2019 weakly excites
- L26:1178 weakly excites
- L27:2382 weakly suppresses
- L26:935 weakly suppresses
- L27:2603 weakly excites
- L26:1038 weakly suppresses
- L26:1296 weakly suppresses word-start 'mult' fragment
- L27:940 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
- c
- c
- (c
- *c
- )c
- =c
- #c
- [c
- ,c
- $c
Suppresses
- Consolid
- 铉
- 党组书记
- работка
- ไกล
- ┬
- ฉบ
- раф
- lửa
- нд