'least' in 'at least' hedges
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. 11 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
Fires on the word 'least' when it follows 'at' in phrases like 'at least not yet' or 'or at least the elder males'. These are hedges that qualify a claim. Turning it up did not produce hedging in the writing, so the effect on output is unclear.
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.
-
new research project at MIT called Sparsh is aiming to fix that oversight.⏎Sparsh (the Hindi word for "touch") isn't an app, at least not in the way we generally use the word. It's a tool that's
-
’s no way around the laws of thermodynamics, at least not yet. Quality of food is an important issue as well, don’t get
-
homeschool. Along the way, we have purposely sought out resources and opportunities to develop fluency in a second language.⏎Finding materials for Norwegian is not easy (at least where I live) so I was very excited to discover the New Amigos board
-
the Industrial Revolution (XVIII century) and the golden age of British education throughout the Victorian era (XIX century).⏎With the information that training, and particularly science schooling, is an endeavor that relies not solely on what occurs in the classroom
-
regardless of whether or not they will be able to fill that extra time with something profound. In a state of nature, people are less concerned with death (at least as an abstract entity, obviously the predator lurking in the bushes is a more real
-
tool that writes your application - your code is a set of instructions to the compiler about what you want it to do. Since compilers aren't very creative (at least we hope) you have to be very precise, and the compiler tends to be
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: the difference between ... ... ... This is a great article. ... ... ... ... It supports the view that ... ... ... Based on the text message ... ... This is peer-to-peer (PMMSO) ... ... ... ... A
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
(2009). The New Physics for the Twenty-First Century. Cambridge University Press.⏎C. W. Kilmister (1994/2005). Eddington's Search for a Fundamental Theory: A
-
clearly indicates that it belongs to the earliest period at Makkah. Traditions also show that the revelations were suspended for a time, which caused the Holy Prophet (upon whom be Allah's peace) to be deeply distressed and grieved.
-
: Krista Tippett⏎Managing Producer: Kate Moos⏎Associate Producer: Nancy Rosenbaum⏎Associate Producer: Shubha Bala⏎Technical Director/Producer: Chris Heagle⏎Senior Editor: Trent Gilliss⏎Associate Web Developer
-
species as having “orthographic” abilities.⏎In the experiment, pigeons were trained to peck four-letter English words as they came up on a screen, or to instead peck a symbol when a four-letter non-word, such as “
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
- L9:2010 suppresses
- L10:1459 suppresses
- L10:135 excites 'or' in alternative requirements
- L6:1373 weakly excites
- L0:1286 weakly excites
- L10:144 weakly suppresses
- L7:2716 weakly excites
- L10:89 weakly suppresses
- L7:625 weakly excites
- L2:493 weakly suppresses
- L10:322 weakly excites
- L2:2440 weakly suppresses
- L10:2650 weakly suppresses
- L1:2568 weakly excites
- L2:319 weakly suppresses 'of' after 'pattern' or 'patterns'
- L7:428 weakly excites
- L10:2535 weakly excites
- L9:464 weakly suppresses
- L4:108 weakly excites "or" offering a choice in web text
- L10:1124 weakly suppresses
- L5:200 weakly excites
- L9:1624 weakly suppresses
- L1:2890 weakly suppresses
- L8:765 weakly excites
Feeds later layers
- L16:1124 suppresses
- L18:2557 weakly excites
- L14:2276 weakly excites opening bracket after a term or number
- L14:2820 weakly excites "and" or "or" linking two counted items
- L25:2471 weakly suppresses
- L15:1501 weakly excites "and" joining two parts of a name or phrase
- L15:2010 weakly suppresses
- L12:969 weakly suppresses
- L12:1452 weakly excites
- L15:2993 weakly suppresses
- L14:2784 weakly excites
- L14:1256 weakly suppresses
- L16:215 weakly excites
- L20:1750 weakly suppresses
- L14:2595 weakly excites
- L17:727 weakly excites
- L21:750 weakly suppresses
- L17:1318 weakly excites
- L15:1189 weakly suppresses
- L18:1133 weakly excites
- L16:2049 weakly excites
- L14:2972 weakly suppresses
- L13:2923 weakly excites
- L16:1454 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
- apest
- เฉพาะ
- ippers
- 鲘
- antic
- ArrayOf
- AsyncResult
- any
- apol
- unch
Suppresses
- oid
- ishes
- N
- aids
- 憾
- オ
- oi
- ftime
- 各
- alphabet