'a' before a time span
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. 24 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 'a' when it comes right before an amount of time, like 'a few days' or 'a short amount of time', and it expects a word such as days or weeks to follow. Turning it up made the model write lists of time words like days, weeks, months and hours.
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.
-
after a preposition to refer again to something that has already been mentioned. See the full definition for<|endoftext|>I have recently been re-reading two books I purchased a few years ago, when I was still teaching languages traditionally. The books are:⏎
-
had multiple FBI sources saying they were 99% sure that Hillary Clinton’s email server had been hacked by foreign intelligence agencies (the claim was retracted a few days later), that was surely Fake News. The content built on arguably real
-
led to him receiving many scars and more importantly brain damages, landing him into a truly deep coma. Fascinatingly, he was able to recover in less than a year of his predicted “revival”. The key factor here is progesterone
-
Women’s Health Watch pointed out, the fact your urine smells stronger is not necessarily a cause for concern.⏎In fact, it’s almost always temporary and disappears in a short amount of time. Despite this, it shouldn’t be ignored that it can
-
hit your head or have a whip-lash type injury. No imaging can identify a concussion and it is purely assessed on signs and symptoms that occur immediately or within a few days following.⏎80-90% of people with a concussion make
-
to aggressively fight for a fair and just settlement.⏎What does a head injury affect quality of life?⏎Because the range of severity is so great with some symptoms lasting a few days to a week, or they could last years or even a lifetime.
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 number of my children's next Nerd's gallery is just one more than the number of my next Nerd's gallery system. Let's make a more precise system now system. Let's make a system for each one of them,
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
water.⏎Feed tropical houseplants once a month during spring and summer, using a water-soluble houseplant fertilizer. During fall and winter, water only once every other month. Or if you prefer, you can fertilize the plant with a
-
cancer would return.⏎They looked at 6,847 women with the ER-positive version of the disease, half of whom were given tamoxifen for the standard five year period, and half who continued with the drug until the 1
-
If I told you there was a crisis that, in mortality terms, was the equivalent of a three 9/11s every week for the last 7 years⏎, would you know which one I'm talking about?⏎
-
digital gold or halving if you’ve been listening close enough. More importantly, you may have even heard about how Bitcoin or BTC prices have been hitting record highs these few days.⏎Also See: The Most Secure Hardware Wallet⏎So, you may
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
- L21:1284 strongly excites
- L19:3065 strongly excites 'for' after a span of time
- L22:1074 suppresses
- L20:1355 suppresses
- L19:1917 excites
- L21:1797 suppresses
- L20:1155 excites
- L22:1699 excites
- L23:1514 suppresses
- L19:2563 excites
- L17:2544 suppresses
- L22:2704 excites
- L14:2842 suppresses
- L21:1729 excites
- L19:1712 excites
- L14:2425 suppresses
- L21:2852 suppresses spot where a number or 'time' comes next
- L7:1114 suppresses
- L20:696 excites
- L12:2377 excites
- L7:2522 suppresses
- L17:1512 excites
- L19:2872 excites
- L17:2964 suppresses
Feeds later layers
- L27:1131 excites
- L25:2993 excites
- L26:284 excites
- L27:1105 excites
- L27:1969 excites
- L26:2126 excites
- L25:565 excites
- L27:1939 excites
- L27:1743 excites
- L25:1286 excites
- L25:2035 weakly excites
- L25:437 weakly excites
- L25:2493 weakly excites
- L26:856 weakly excites
- L25:2817 weakly excites
- L26:1448 weakly excites
- L25:1377 weakly excites
- L27:423 weakly excites
- L26:1445 weakly excites
- L27:1233 weakly excites
- L26:2225 weakly excites
- L26:2769 weakly excites the 'inter' word beginning
- L26:2543 weakly excites
- L27:2760 weakly excites 'the' after 'between' in age ranges
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
- months
- days
- weeks
- hours
- years
- days
- weeks
- Months
- months
- hours
Suppresses
- 第三个
- MatrixMode
- listOf
- istine
- 饧
- emouth
- nex
- knife
- 石灰
- -profile