the blank slot before a year
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. 18 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 spot just before a year number, such as the gap in 'in [blank]2015' or 'the year [blank]2008'. It is the point where the model is about to write a year. Turning it down made the model keep writing years, and the negative side is on numbers and words in unit or measurement contexts.
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.
-
be aware of your weaknesses and thus exploit them more effectively than an outside agent might be able to.⏎This chart shows the type of people that committed cyberattacks in 2015.<|endoftext|>NPR’s Fresh Air book review mentions Georgia Ch
-
, multi-scale nested watersheds were delineated and the relationships between two representative water quality indexes and agricultural land use intensity were assessed and quantified for the year 2000 using multi-scale regression analysis. The results show that the log
-
district come from a variety of sources. For income taxes paid, we used the most recent dataset released by the IRS, which breaks down tax information from the year 2007 by zip code. We then link zip codes to congressional districts
-
grown, its headquarters have expanded into the surrounding area. In 1917, an administration building was built on the block east of the temple and in 1972, the twenty-eight story LDS Church Office Building, which was
-
. Studies show that though only 31.4 per cent of the adult populace had teeth rotting in the year 1994; the year 2008 saw this percentage increase to a shocking 39.9
-
ry disease characteristics, and the clinical parameters.⏎The investigators scoured the Pubmed, EMBASE, and CINAHL databases from their inception to February of 2018 to identify articles on Fabry disease and white matter lesions examined
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 30000000000000000000000000000000000000000000
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
phenomena, directly and indirectly, constitute active transport of matter due to an applied electric current. The transport is measured in units of chemical flux, commonly µmol/cm2h. Iontophoresis has experimental, therapeutic and diagnostic applications.⏎I
-
Glenn Fowler, M.D., M.P.H., of the Makerere University–Johns Hopkins University Research Collaboration in Kampala, Uganda, the study team expects results in five to six years.⏎The HIV-infected women eligible to participate in PROM
-
are released. The eggs are semi-buoyant, causing them to be dispersed widely by the water current. The eggs develop rapidly, hatching into larvae within 24 to 48 hours after fertilisation (8).<|endoftext|>Peter
-
is a communicable disease, which is caused by a recently revealed coronavirus. Those who are infected with the virus will experience mild to modest respiratory sickness. They can recover from the illness without the need for special treatment. The older population is more vulnerable
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
- L17:2964 strongly suppresses
- L17:2750 strongly suppresses
- L17:2760 strongly excites
- L17:2544 strongly suppresses
- L12:2593 strongly excites
- L12:2691 strongly suppresses
- L14:2201 excites
- L16:1815 excites
- L15:1520 excites
- L17:1041 excites
- L12:2338 excites
- L15:2447 suppresses
- L14:2939 excites
- L14:1088 excites
- L10:637 suppresses
- L14:1778 suppresses
- L15:1825 excites
- L10:986 excites
- L11:1352 excites
- L11:2530 excites
- L16:1050 suppresses
- L16:1668 suppresses
- L16:1930 excites
- L15:2851 suppresses
Feeds later layers
- L22:1074 strongly suppresses
- L25:565 suppresses
- L21:741 suppresses
- L19:245 excites
- L20:1680 suppresses
- L20:2472 suppresses
- L21:848 suppresses
- L20:2347 suppresses
- L22:2371 suppresses
- L22:2932 suppresses
- L19:2242 suppresses
- L20:2792 suppresses
- L26:284 suppresses
- L24:906 suppresses
- L21:1729 suppresses
- L22:1157 suppresses
- L20:1335 suppresses
- L22:2445 suppresses
- L23:1919 suppresses
- L27:934 suppresses
- L19:1720 suppresses
- L25:2440 suppresses
- L23:1337 suppresses
- L19:2474 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
- ambi
- ButtonTitles
- 的手
- ragments
- enticated
- 幂
- getState
- 貌
- мот
- 一脚
Suppresses
- time
- 时间
- 时间
- 的时间
- stime
- 时间为
- etime
- Time
- _Time
- $time