first digit of a year like 2020
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 responds to the first digit of a year, mostly the "2" in 2000s dates such as 2020 or 2017. It is weaker on the leading digit of other measurements. When turned up, the model's output changed topic and became more ordinary prose, while turning it down made the output a string of zeros.
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.
-
change. Even by California standards; it’s ambitious, calling for a 36% reduction in greenhouse gas emissions from 2005 levels by 2020 (statewide emissions decreased 5.3% between 2
-
Solution : Proposed by UN⏎- India’s take on it⏎Reasons of Conflict:⏎- The renewed violence in the region started on May 6, 2021 when Palestinians protested against an anticipated decision of the Israeli Supreme Court over
-
3(3), pages 315-320.⏎- Angus Deaton & Olivier Dupriez, 2011.⏎"Purchasing Power Parity Exchange Rates for the Global Poor
-
the issue by modifying the ALA<|endoftext|>Note: I published a previous version of this article on The Change Blog (now called Possibility of Change) in 2011.⏎What is life? Is it the state of physical existence,
-
sur l’évolution des charges et produits de l’Assurance Maladie au titre de 2014 (loi du 13 août 20
-
requirements of the PCI Security Standards Council. Outdated browsers do not support the requirements of modern web applications and pose significant security risks. After May 1, 2017, this browser will no longer be supported and you may experience issues
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 0.50000000000000000000000000000000000000000000
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
36 short tons)|⏎|Length:||205 ft (62.48 m)o/a|⏎|Beam:||33 ft (10.06 m)|⏎|Draught:||11.
-
New York factory, which was a one-piece brass measuring 33.5 meters (109’) tall. This flagpole is still the tallest in service at An
-
500 UTC (11:00 a.m), Humberto had weakened to tropical storm status, with maximum sustained winds of 65 mph (105 km/h).⏎On September 14, Humberto had reached post
-
gent plate margin between the African Plate and the Eurasian Plate. It is the tallest active volcano in Europe, currently standing 3,329 m (10,922 ft) high, though this varies with summit eruptions
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
- L24:2703 strongly suppresses
- L23:2761 excites
- L24:2118 excites
- L24:2879 suppresses
- L22:1187 suppresses
- L22:1505 excites
- L23:1295 suppresses
- L24:1676 excites zeros in large round numbers
- L23:510 suppresses
- L23:929 excites
- L24:430 suppresses
- L24:1777 suppresses
- L23:2629 suppresses
- L22:909 excites
- L23:2924 suppresses
- L23:2222 excites
- L23:615 excites
- L23:950 suppresses
- L21:2233 excites
- L24:2757 excites
- L24:2618 suppresses
- L22:1095 suppresses
- L22:434 excites
- L21:242 suppresses 'very' before an adjective
Feeds later layers
- L27:1153 strongly suppresses
- L26:413 strongly suppresses
- L27:862 suppresses
- L27:153 excites the spot just before a historical number
- L26:848 suppresses
- L27:1202 suppresses
- L27:396 suppresses
- L26:682 excites
- L26:598 excites spot just before a year or century
- L26:1126 suppresses the digit 1 in powers of ten
- L27:1522 weakly excites
- L27:328 weakly excites
- L27:352 weakly excites
- L26:834 weakly excites
- L27:992 weakly excites
- L27:180 weakly suppresses
- L26:1824 weakly suppresses
- L26:1521 weakly suppresses
- L27:1111 weakly suppresses
- L26:724 weakly excites
- L27:3011 weakly excites
- L27:902 weakly excites
- L26:2255 weakly excites
- L27:189 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
- hare
- piel
- ingers
- ethyl
- EntityType
- yro
- urance
- 界第一
- eterminate
- igar
Suppresses
- 0
- 0
- ₀
- lify
- tas
- 迈进
- 零
- zero
- newline
- 动力