"not" or "instead" contrasting a correction
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. 6 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 words like "not" when they add a correction or contrast, such as "not even" or "not X, not Y". It switches off on "instead" and "but" in contrast phrases. Turning it up gave repetitive or unrelated text, so the steering doesn't clearly reflect this.
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.
-
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 supposed
-
Waziristan. Consequently, it became a sanctuary for fundamentalist terrorists who hold outdated ultraconservative views that are particularly detrimental for women.⏎But nothing — not even a conservative culture or fear of the Taliban<|endoftext|>This preview shows pages 1
-
, but the more pressure in the game, the more he seemed to “forget” how to throw.⏎Knoblauch is not the first and probably not the last player this has happened to. Tennis players ‘forget’ how to toss
-
them to fulfil their promises no matter what. We are also given access to social security and food parcels from the government so we wouldn’t starve, at least not starve enough to find a job. Those that manage to get a job,
-
, it is in a circle, at most perhaps in a widening circle. Man today is not wiser than the ancient seers and sages and thinkers, not more spiritual than the great seekers of old, the first mighty mystics, not
-
Rabbi Emil Fackenheim in his book, "God's Presence in History." "Nevertheless, not messengers, not angels, not intermediaries, but God himself acts in human history."⏎God hears
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: what is the value of the question is to be given in a diagram of a cube (a regular hexagon) and a rectangle, and a square, and a circle, and a square of a circle, and a square of a circle
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
is a semiconductor device that can be programmed to perform various logic functions. Unlike a microcontroller, an FPGA is not designed to execute a specific set of instructions. Instead, developers program the FPGA to perform specific logic functions by configuring its hardware components.
-
example, your tomato ketchup or mayonnaise⏎- Stop using plastic drinking straws and instead try metal, paper, or even wheat straws⏎- Swap tampons and
-
says, “This he said, not that he cared for the poor, but because he was a thief, and had the money box; and he used to
-
any decision, then you will be well-informed of all the consequences of your decision-making. Critical thinking doesn’t mean that your decision will never be wrong. Instead, it increases your chances of making good decisions. Although if the outcome is not
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
- L5:45 weakly suppresses
- L5:2082 weakly suppresses
- L3:807 weakly suppresses comma inside dates and years
- L5:57 weakly suppresses
- L1:2539 weakly suppresses
- L5:60 weakly excites
- L1:2028 weakly suppresses
- L3:1468 weakly excites
- L3:1377 weakly suppresses
- L0:2285 weakly excites
- L4:248 weakly suppresses
- L5:96 weakly suppresses
- L2:2027 weakly suppresses
- L3:1382 weakly suppresses
- L3:964 weakly excites
- L3:2126 weakly excites
- L2:851 weakly suppresses
- L4:1901 weakly suppresses
- L0:2946 weakly excites
- L5:1326 weakly suppresses
- L3:1883 weakly excites
- L2:1194 weakly suppresses
- L5:169 weakly excites
- L1:2681 weakly suppresses
Feeds later layers
- L8:1784 suppresses
- L8:1876 suppresses
- L14:2166 suppresses
- L7:678 excites
- L8:1011 suppresses
- L7:1846 suppresses
- L11:1861 suppresses
- L12:1071 suppresses
- L14:2811 excites
- L7:2256 suppresses 'to' after come or bring
- L14:2713 excites
- L18:1553 suppresses
- L9:2343 suppresses
- L27:1088 weakly suppresses
- L7:1764 weakly excites
- L9:1046 weakly suppresses
- L13:220 weakly suppresses
- L7:2806 weakly excites
- L9:1133 weakly suppresses
- L9:1472 weakly excites
- L11:2662 weakly suppresses
- L9:2203 weakly excites
- L19:1811 weakly suppresses
- L13:92 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
- Late
- alyzed
- ��
- イベ
- 孽
- Tucker
- ngth
- 会有
- nier
- レビュー
Suppresses
- instead
- instead
- rather
- Instead
- Instead
- .Ptr
- 呢
- (d
- downright
- ignum